linux/mm/vmscan.c
<<
>>
Prefs
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   4 *
   5 *  Swap reorganised 29.12.95, Stephen Tweedie.
   6 *  kswapd added: 7.1.96  sct
   7 *  Removed kswapd_ctl limits, and swap out as many pages as needed
   8 *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
   9 *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
  10 *  Multiqueue VM started 5.8.00, Rik van Riel.
  11 */
  12
  13#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  14
  15#include <linux/mm.h>
  16#include <linux/sched/mm.h>
  17#include <linux/module.h>
  18#include <linux/gfp.h>
  19#include <linux/kernel_stat.h>
  20#include <linux/swap.h>
  21#include <linux/pagemap.h>
  22#include <linux/init.h>
  23#include <linux/highmem.h>
  24#include <linux/vmpressure.h>
  25#include <linux/vmstat.h>
  26#include <linux/file.h>
  27#include <linux/writeback.h>
  28#include <linux/blkdev.h>
  29#include <linux/buffer_head.h>  /* for buffer_heads_over_limit */
  30#include <linux/mm_inline.h>
  31#include <linux/backing-dev.h>
  32#include <linux/rmap.h>
  33#include <linux/topology.h>
  34#include <linux/cpu.h>
  35#include <linux/cpuset.h>
  36#include <linux/compaction.h>
  37#include <linux/notifier.h>
  38#include <linux/rwsem.h>
  39#include <linux/delay.h>
  40#include <linux/kthread.h>
  41#include <linux/freezer.h>
  42#include <linux/memcontrol.h>
  43#include <linux/migrate.h>
  44#include <linux/delayacct.h>
  45#include <linux/sysctl.h>
  46#include <linux/oom.h>
  47#include <linux/pagevec.h>
  48#include <linux/prefetch.h>
  49#include <linux/printk.h>
  50#include <linux/dax.h>
  51#include <linux/psi.h>
  52
  53#include <asm/tlbflush.h>
  54#include <asm/div64.h>
  55
  56#include <linux/swapops.h>
  57#include <linux/balloon_compaction.h>
  58#include <linux/sched/sysctl.h>
  59
  60#include "internal.h"
  61#include "swap.h"
  62
  63#define CREATE_TRACE_POINTS
  64#include <trace/events/vmscan.h>
  65
  66struct scan_control {
  67        /* How many pages shrink_list() should reclaim */
  68        unsigned long nr_to_reclaim;
  69
  70        /*
  71         * Nodemask of nodes allowed by the caller. If NULL, all nodes
  72         * are scanned.
  73         */
  74        nodemask_t      *nodemask;
  75
  76        /*
  77         * The memory cgroup that hit its limit and as a result is the
  78         * primary target of this reclaim invocation.
  79         */
  80        struct mem_cgroup *target_mem_cgroup;
  81
  82        /*
  83         * Scan pressure balancing between anon and file LRUs
  84         */
  85        unsigned long   anon_cost;
  86        unsigned long   file_cost;
  87
  88        /* Can active pages be deactivated as part of reclaim? */
  89#define DEACTIVATE_ANON 1
  90#define DEACTIVATE_FILE 2
  91        unsigned int may_deactivate:2;
  92        unsigned int force_deactivate:1;
  93        unsigned int skipped_deactivate:1;
  94
  95        /* Writepage batching in laptop mode; RECLAIM_WRITE */
  96        unsigned int may_writepage:1;
  97
  98        /* Can mapped pages be reclaimed? */
  99        unsigned int may_unmap:1;
 100
 101        /* Can pages be swapped as part of reclaim? */
 102        unsigned int may_swap:1;
 103
 104        /* Proactive reclaim invoked by userspace through memory.reclaim */
 105        unsigned int proactive:1;
 106
 107        /*
 108         * Cgroup memory below memory.low is protected as long as we
 109         * don't threaten to OOM. If any cgroup is reclaimed at
 110         * reduced force or passed over entirely due to its memory.low
 111         * setting (memcg_low_skipped), and nothing is reclaimed as a
 112         * result, then go back for one more cycle that reclaims the protected
 113         * memory (memcg_low_reclaim) to avert OOM.
 114         */
 115        unsigned int memcg_low_reclaim:1;
 116        unsigned int memcg_low_skipped:1;
 117
 118        unsigned int hibernation_mode:1;
 119
 120        /* One of the zones is ready for compaction */
 121        unsigned int compaction_ready:1;
 122
 123        /* There is easily reclaimable cold cache in the current node */
 124        unsigned int cache_trim_mode:1;
 125
 126        /* The file pages on the current node are dangerously low */
 127        unsigned int file_is_tiny:1;
 128
 129        /* Always discard instead of demoting to lower tier memory */
 130        unsigned int no_demotion:1;
 131
 132        /* Allocation order */
 133        s8 order;
 134
 135        /* Scan (total_size >> priority) pages at once */
 136        s8 priority;
 137
 138        /* The highest zone to isolate pages for reclaim from */
 139        s8 reclaim_idx;
 140
 141        /* This context's GFP mask */
 142        gfp_t gfp_mask;
 143
 144        /* Incremented by the number of inactive pages that were scanned */
 145        unsigned long nr_scanned;
 146
 147        /* Number of pages freed so far during a call to shrink_zones() */
 148        unsigned long nr_reclaimed;
 149
 150        struct {
 151                unsigned int dirty;
 152                unsigned int unqueued_dirty;
 153                unsigned int congested;
 154                unsigned int writeback;
 155                unsigned int immediate;
 156                unsigned int file_taken;
 157                unsigned int taken;
 158        } nr;
 159
 160        /* for recording the reclaimed slab by now */
 161        struct reclaim_state reclaim_state;
 162};
 163
 164#ifdef ARCH_HAS_PREFETCHW
 165#define prefetchw_prev_lru_folio(_folio, _base, _field)                 \
 166        do {                                                            \
 167                if ((_folio)->lru.prev != _base) {                      \
 168                        struct folio *prev;                             \
 169                                                                        \
 170                        prev = lru_to_folio(&(_folio->lru));            \
 171                        prefetchw(&prev->_field);                       \
 172                }                                                       \
 173        } while (0)
 174#else
 175#define prefetchw_prev_lru_folio(_folio, _base, _field) do { } while (0)
 176#endif
 177
 178/*
 179 * From 0 .. 200.  Higher means more swappy.
 180 */
 181int vm_swappiness = 60;
 182
 183static void set_task_reclaim_state(struct task_struct *task,
 184                                   struct reclaim_state *rs)
 185{
 186        /* Check for an overwrite */
 187        WARN_ON_ONCE(rs && task->reclaim_state);
 188
 189        /* Check for the nulling of an already-nulled member */
 190        WARN_ON_ONCE(!rs && !task->reclaim_state);
 191
 192        task->reclaim_state = rs;
 193}
 194
 195LIST_HEAD(shrinker_list);
 196DECLARE_RWSEM(shrinker_rwsem);
 197
 198#ifdef CONFIG_MEMCG
 199static int shrinker_nr_max;
 200
 201/* The shrinker_info is expanded in a batch of BITS_PER_LONG */
 202static inline int shrinker_map_size(int nr_items)
 203{
 204        return (DIV_ROUND_UP(nr_items, BITS_PER_LONG) * sizeof(unsigned long));
 205}
 206
 207static inline int shrinker_defer_size(int nr_items)
 208{
 209        return (round_up(nr_items, BITS_PER_LONG) * sizeof(atomic_long_t));
 210}
 211
 212static struct shrinker_info *shrinker_info_protected(struct mem_cgroup *memcg,
 213                                                     int nid)
 214{
 215        return rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_info,
 216                                         lockdep_is_held(&shrinker_rwsem));
 217}
 218
 219static int expand_one_shrinker_info(struct mem_cgroup *memcg,
 220                                    int map_size, int defer_size,
 221                                    int old_map_size, int old_defer_size)
 222{
 223        struct shrinker_info *new, *old;
 224        struct mem_cgroup_per_node *pn;
 225        int nid;
 226        int size = map_size + defer_size;
 227
 228        for_each_node(nid) {
 229                pn = memcg->nodeinfo[nid];
 230                old = shrinker_info_protected(memcg, nid);
 231                /* Not yet online memcg */
 232                if (!old)
 233                        return 0;
 234
 235                new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
 236                if (!new)
 237                        return -ENOMEM;
 238
 239                new->nr_deferred = (atomic_long_t *)(new + 1);
 240                new->map = (void *)new->nr_deferred + defer_size;
 241
 242                /* map: set all old bits, clear all new bits */
 243                memset(new->map, (int)0xff, old_map_size);
 244                memset((void *)new->map + old_map_size, 0, map_size - old_map_size);
 245                /* nr_deferred: copy old values, clear all new values */
 246                memcpy(new->nr_deferred, old->nr_deferred, old_defer_size);
 247                memset((void *)new->nr_deferred + old_defer_size, 0,
 248                       defer_size - old_defer_size);
 249
 250                rcu_assign_pointer(pn->shrinker_info, new);
 251                kvfree_rcu(old, rcu);
 252        }
 253
 254        return 0;
 255}
 256
 257void free_shrinker_info(struct mem_cgroup *memcg)
 258{
 259        struct mem_cgroup_per_node *pn;
 260        struct shrinker_info *info;
 261        int nid;
 262
 263        for_each_node(nid) {
 264                pn = memcg->nodeinfo[nid];
 265                info = rcu_dereference_protected(pn->shrinker_info, true);
 266                kvfree(info);
 267                rcu_assign_pointer(pn->shrinker_info, NULL);
 268        }
 269}
 270
 271int alloc_shrinker_info(struct mem_cgroup *memcg)
 272{
 273        struct shrinker_info *info;
 274        int nid, size, ret = 0;
 275        int map_size, defer_size = 0;
 276
 277        down_write(&shrinker_rwsem);
 278        map_size = shrinker_map_size(shrinker_nr_max);
 279        defer_size = shrinker_defer_size(shrinker_nr_max);
 280        size = map_size + defer_size;
 281        for_each_node(nid) {
 282                info = kvzalloc_node(sizeof(*info) + size, GFP_KERNEL, nid);
 283                if (!info) {
 284                        free_shrinker_info(memcg);
 285                        ret = -ENOMEM;
 286                        break;
 287                }
 288                info->nr_deferred = (atomic_long_t *)(info + 1);
 289                info->map = (void *)info->nr_deferred + defer_size;
 290                rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_info, info);
 291        }
 292        up_write(&shrinker_rwsem);
 293
 294        return ret;
 295}
 296
 297static inline bool need_expand(int nr_max)
 298{
 299        return round_up(nr_max, BITS_PER_LONG) >
 300               round_up(shrinker_nr_max, BITS_PER_LONG);
 301}
 302
 303static int expand_shrinker_info(int new_id)
 304{
 305        int ret = 0;
 306        int new_nr_max = new_id + 1;
 307        int map_size, defer_size = 0;
 308        int old_map_size, old_defer_size = 0;
 309        struct mem_cgroup *memcg;
 310
 311        if (!need_expand(new_nr_max))
 312                goto out;
 313
 314        if (!root_mem_cgroup)
 315                goto out;
 316
 317        lockdep_assert_held(&shrinker_rwsem);
 318
 319        map_size = shrinker_map_size(new_nr_max);
 320        defer_size = shrinker_defer_size(new_nr_max);
 321        old_map_size = shrinker_map_size(shrinker_nr_max);
 322        old_defer_size = shrinker_defer_size(shrinker_nr_max);
 323
 324        memcg = mem_cgroup_iter(NULL, NULL, NULL);
 325        do {
 326                ret = expand_one_shrinker_info(memcg, map_size, defer_size,
 327                                               old_map_size, old_defer_size);
 328                if (ret) {
 329                        mem_cgroup_iter_break(NULL, memcg);
 330                        goto out;
 331                }
 332        } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
 333out:
 334        if (!ret)
 335                shrinker_nr_max = new_nr_max;
 336
 337        return ret;
 338}
 339
 340void set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
 341{
 342        if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
 343                struct shrinker_info *info;
 344
 345                rcu_read_lock();
 346                info = rcu_dereference(memcg->nodeinfo[nid]->shrinker_info);
 347                /* Pairs with smp mb in shrink_slab() */
 348                smp_mb__before_atomic();
 349                set_bit(shrinker_id, info->map);
 350                rcu_read_unlock();
 351        }
 352}
 353
 354static DEFINE_IDR(shrinker_idr);
 355
 356static int prealloc_memcg_shrinker(struct shrinker *shrinker)
 357{
 358        int id, ret = -ENOMEM;
 359
 360        if (mem_cgroup_disabled())
 361                return -ENOSYS;
 362
 363        down_write(&shrinker_rwsem);
 364        /* This may call shrinker, so it must use down_read_trylock() */
 365        id = idr_alloc(&shrinker_idr, shrinker, 0, 0, GFP_KERNEL);
 366        if (id < 0)
 367                goto unlock;
 368
 369        if (id >= shrinker_nr_max) {
 370                if (expand_shrinker_info(id)) {
 371                        idr_remove(&shrinker_idr, id);
 372                        goto unlock;
 373                }
 374        }
 375        shrinker->id = id;
 376        ret = 0;
 377unlock:
 378        up_write(&shrinker_rwsem);
 379        return ret;
 380}
 381
 382static void unregister_memcg_shrinker(struct shrinker *shrinker)
 383{
 384        int id = shrinker->id;
 385
 386        BUG_ON(id < 0);
 387
 388        lockdep_assert_held(&shrinker_rwsem);
 389
 390        idr_remove(&shrinker_idr, id);
 391}
 392
 393static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker,
 394                                   struct mem_cgroup *memcg)
 395{
 396        struct shrinker_info *info;
 397
 398        info = shrinker_info_protected(memcg, nid);
 399        return atomic_long_xchg(&info->nr_deferred[shrinker->id], 0);
 400}
 401
 402static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker,
 403                                  struct mem_cgroup *memcg)
 404{
 405        struct shrinker_info *info;
 406
 407        info = shrinker_info_protected(memcg, nid);
 408        return atomic_long_add_return(nr, &info->nr_deferred[shrinker->id]);
 409}
 410
 411void reparent_shrinker_deferred(struct mem_cgroup *memcg)
 412{
 413        int i, nid;
 414        long nr;
 415        struct mem_cgroup *parent;
 416        struct shrinker_info *child_info, *parent_info;
 417
 418        parent = parent_mem_cgroup(memcg);
 419        if (!parent)
 420                parent = root_mem_cgroup;
 421
 422        /* Prevent from concurrent shrinker_info expand */
 423        down_read(&shrinker_rwsem);
 424        for_each_node(nid) {
 425                child_info = shrinker_info_protected(memcg, nid);
 426                parent_info = shrinker_info_protected(parent, nid);
 427                for (i = 0; i < shrinker_nr_max; i++) {
 428                        nr = atomic_long_read(&child_info->nr_deferred[i]);
 429                        atomic_long_add(nr, &parent_info->nr_deferred[i]);
 430                }
 431        }
 432        up_read(&shrinker_rwsem);
 433}
 434
 435static bool cgroup_reclaim(struct scan_control *sc)
 436{
 437        return sc->target_mem_cgroup;
 438}
 439
 440/**
 441 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
 442 * @sc: scan_control in question
 443 *
 444 * The normal page dirty throttling mechanism in balance_dirty_pages() is
 445 * completely broken with the legacy memcg and direct stalling in
 446 * shrink_page_list() is used for throttling instead, which lacks all the
 447 * niceties such as fairness, adaptive pausing, bandwidth proportional
 448 * allocation and configurability.
 449 *
 450 * This function tests whether the vmscan currently in progress can assume
 451 * that the normal dirty throttling mechanism is operational.
 452 */
 453static bool writeback_throttling_sane(struct scan_control *sc)
 454{
 455        if (!cgroup_reclaim(sc))
 456                return true;
 457#ifdef CONFIG_CGROUP_WRITEBACK
 458        if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
 459                return true;
 460#endif
 461        return false;
 462}
 463#else
 464static int prealloc_memcg_shrinker(struct shrinker *shrinker)
 465{
 466        return -ENOSYS;
 467}
 468
 469static void unregister_memcg_shrinker(struct shrinker *shrinker)
 470{
 471}
 472
 473static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker,
 474                                   struct mem_cgroup *memcg)
 475{
 476        return 0;
 477}
 478
 479static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker,
 480                                  struct mem_cgroup *memcg)
 481{
 482        return 0;
 483}
 484
 485static bool cgroup_reclaim(struct scan_control *sc)
 486{
 487        return false;
 488}
 489
 490static bool writeback_throttling_sane(struct scan_control *sc)
 491{
 492        return true;
 493}
 494#endif
 495
 496static long xchg_nr_deferred(struct shrinker *shrinker,
 497                             struct shrink_control *sc)
 498{
 499        int nid = sc->nid;
 500
 501        if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
 502                nid = 0;
 503
 504        if (sc->memcg &&
 505            (shrinker->flags & SHRINKER_MEMCG_AWARE))
 506                return xchg_nr_deferred_memcg(nid, shrinker,
 507                                              sc->memcg);
 508
 509        return atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
 510}
 511
 512
 513static long add_nr_deferred(long nr, struct shrinker *shrinker,
 514                            struct shrink_control *sc)
 515{
 516        int nid = sc->nid;
 517
 518        if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
 519                nid = 0;
 520
 521        if (sc->memcg &&
 522            (shrinker->flags & SHRINKER_MEMCG_AWARE))
 523                return add_nr_deferred_memcg(nr, nid, shrinker,
 524                                             sc->memcg);
 525
 526        return atomic_long_add_return(nr, &shrinker->nr_deferred[nid]);
 527}
 528
 529static bool can_demote(int nid, struct scan_control *sc)
 530{
 531        if (!numa_demotion_enabled)
 532                return false;
 533        if (sc && sc->no_demotion)
 534                return false;
 535        if (next_demotion_node(nid) == NUMA_NO_NODE)
 536                return false;
 537
 538        return true;
 539}
 540
 541static inline bool can_reclaim_anon_pages(struct mem_cgroup *memcg,
 542                                          int nid,
 543                                          struct scan_control *sc)
 544{
 545        if (memcg == NULL) {
 546                /*
 547                 * For non-memcg reclaim, is there
 548                 * space in any swap device?
 549                 */
 550                if (get_nr_swap_pages() > 0)
 551                        return true;
 552        } else {
 553                /* Is the memcg below its swap limit? */
 554                if (mem_cgroup_get_nr_swap_pages(memcg) > 0)
 555                        return true;
 556        }
 557
 558        /*
 559         * The page can not be swapped.
 560         *
 561         * Can it be reclaimed from this node via demotion?
 562         */
 563        return can_demote(nid, sc);
 564}
 565
 566/*
 567 * This misses isolated pages which are not accounted for to save counters.
 568 * As the data only determines if reclaim or compaction continues, it is
 569 * not expected that isolated pages will be a dominating factor.
 570 */
 571unsigned long zone_reclaimable_pages(struct zone *zone)
 572{
 573        unsigned long nr;
 574
 575        nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
 576                zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
 577        if (can_reclaim_anon_pages(NULL, zone_to_nid(zone), NULL))
 578                nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
 579                        zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
 580
 581        return nr;
 582}
 583
 584/**
 585 * lruvec_lru_size -  Returns the number of pages on the given LRU list.
 586 * @lruvec: lru vector
 587 * @lru: lru to use
 588 * @zone_idx: zones to consider (use MAX_NR_ZONES - 1 for the whole LRU list)
 589 */
 590static unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru,
 591                                     int zone_idx)
 592{
 593        unsigned long size = 0;
 594        int zid;
 595
 596        for (zid = 0; zid <= zone_idx; zid++) {
 597                struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
 598
 599                if (!managed_zone(zone))
 600                        continue;
 601
 602                if (!mem_cgroup_disabled())
 603                        size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
 604                else
 605                        size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
 606        }
 607        return size;
 608}
 609
 610/*
 611 * Add a shrinker callback to be called from the vm.
 612 */
 613static int __prealloc_shrinker(struct shrinker *shrinker)
 614{
 615        unsigned int size;
 616        int err;
 617
 618        if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
 619                err = prealloc_memcg_shrinker(shrinker);
 620                if (err != -ENOSYS)
 621                        return err;
 622
 623                shrinker->flags &= ~SHRINKER_MEMCG_AWARE;
 624        }
 625
 626        size = sizeof(*shrinker->nr_deferred);
 627        if (shrinker->flags & SHRINKER_NUMA_AWARE)
 628                size *= nr_node_ids;
 629
 630        shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
 631        if (!shrinker->nr_deferred)
 632                return -ENOMEM;
 633
 634        return 0;
 635}
 636
 637#ifdef CONFIG_SHRINKER_DEBUG
 638int prealloc_shrinker(struct shrinker *shrinker, const char *fmt, ...)
 639{
 640        va_list ap;
 641        int err;
 642
 643        va_start(ap, fmt);
 644        shrinker->name = kvasprintf_const(GFP_KERNEL, fmt, ap);
 645        va_end(ap);
 646        if (!shrinker->name)
 647                return -ENOMEM;
 648
 649        err = __prealloc_shrinker(shrinker);
 650        if (err) {
 651                kfree_const(shrinker->name);
 652                shrinker->name = NULL;
 653        }
 654
 655        return err;
 656}
 657#else
 658int prealloc_shrinker(struct shrinker *shrinker, const char *fmt, ...)
 659{
 660        return __prealloc_shrinker(shrinker);
 661}
 662#endif
 663
 664void free_prealloced_shrinker(struct shrinker *shrinker)
 665{
 666#ifdef CONFIG_SHRINKER_DEBUG
 667        kfree_const(shrinker->name);
 668        shrinker->name = NULL;
 669#endif
 670        if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
 671                down_write(&shrinker_rwsem);
 672                unregister_memcg_shrinker(shrinker);
 673                up_write(&shrinker_rwsem);
 674                return;
 675        }
 676
 677        kfree(shrinker->nr_deferred);
 678        shrinker->nr_deferred = NULL;
 679}
 680
 681void register_shrinker_prepared(struct shrinker *shrinker)
 682{
 683        down_write(&shrinker_rwsem);
 684        list_add_tail(&shrinker->list, &shrinker_list);
 685        shrinker->flags |= SHRINKER_REGISTERED;
 686        shrinker_debugfs_add(shrinker);
 687        up_write(&shrinker_rwsem);
 688}
 689
 690static int __register_shrinker(struct shrinker *shrinker)
 691{
 692        int err = __prealloc_shrinker(shrinker);
 693
 694        if (err)
 695                return err;
 696        register_shrinker_prepared(shrinker);
 697        return 0;
 698}
 699
 700#ifdef CONFIG_SHRINKER_DEBUG
 701int register_shrinker(struct shrinker *shrinker, const char *fmt, ...)
 702{
 703        va_list ap;
 704        int err;
 705
 706        va_start(ap, fmt);
 707        shrinker->name = kvasprintf_const(GFP_KERNEL, fmt, ap);
 708        va_end(ap);
 709        if (!shrinker->name)
 710                return -ENOMEM;
 711
 712        err = __register_shrinker(shrinker);
 713        if (err) {
 714                kfree_const(shrinker->name);
 715                shrinker->name = NULL;
 716        }
 717        return err;
 718}
 719#else
 720int register_shrinker(struct shrinker *shrinker, const char *fmt, ...)
 721{
 722        return __register_shrinker(shrinker);
 723}
 724#endif
 725EXPORT_SYMBOL(register_shrinker);
 726
 727/*
 728 * Remove one
 729 */
 730void unregister_shrinker(struct shrinker *shrinker)
 731{
 732        if (!(shrinker->flags & SHRINKER_REGISTERED))
 733                return;
 734
 735        down_write(&shrinker_rwsem);
 736        list_del(&shrinker->list);
 737        shrinker->flags &= ~SHRINKER_REGISTERED;
 738        if (shrinker->flags & SHRINKER_MEMCG_AWARE)
 739                unregister_memcg_shrinker(shrinker);
 740        shrinker_debugfs_remove(shrinker);
 741        up_write(&shrinker_rwsem);
 742
 743        kfree(shrinker->nr_deferred);
 744        shrinker->nr_deferred = NULL;
 745}
 746EXPORT_SYMBOL(unregister_shrinker);
 747
 748/**
 749 * synchronize_shrinkers - Wait for all running shrinkers to complete.
 750 *
 751 * This is equivalent to calling unregister_shrink() and register_shrinker(),
 752 * but atomically and with less overhead. This is useful to guarantee that all
 753 * shrinker invocations have seen an update, before freeing memory, similar to
 754 * rcu.
 755 */
 756void synchronize_shrinkers(void)
 757{
 758        down_write(&shrinker_rwsem);
 759        up_write(&shrinker_rwsem);
 760}
 761EXPORT_SYMBOL(synchronize_shrinkers);
 762
 763#define SHRINK_BATCH 128
 764
 765static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
 766                                    struct shrinker *shrinker, int priority)
 767{
 768        unsigned long freed = 0;
 769        unsigned long long delta;
 770        long total_scan;
 771        long freeable;
 772        long nr;
 773        long new_nr;
 774        long batch_size = shrinker->batch ? shrinker->batch
 775                                          : SHRINK_BATCH;
 776        long scanned = 0, next_deferred;
 777
 778        freeable = shrinker->count_objects(shrinker, shrinkctl);
 779        if (freeable == 0 || freeable == SHRINK_EMPTY)
 780                return freeable;
 781
 782        /*
 783         * copy the current shrinker scan count into a local variable
 784         * and zero it so that other concurrent shrinker invocations
 785         * don't also do this scanning work.
 786         */
 787        nr = xchg_nr_deferred(shrinker, shrinkctl);
 788
 789        if (shrinker->seeks) {
 790                delta = freeable >> priority;
 791                delta *= 4;
 792                do_div(delta, shrinker->seeks);
 793        } else {
 794                /*
 795                 * These objects don't require any IO to create. Trim
 796                 * them aggressively under memory pressure to keep
 797                 * them from causing refetches in the IO caches.
 798                 */
 799                delta = freeable / 2;
 800        }
 801
 802        total_scan = nr >> priority;
 803        total_scan += delta;
 804        total_scan = min(total_scan, (2 * freeable));
 805
 806        trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
 807                                   freeable, delta, total_scan, priority);
 808
 809        /*
 810         * Normally, we should not scan less than batch_size objects in one
 811         * pass to avoid too frequent shrinker calls, but if the slab has less
 812         * than batch_size objects in total and we are really tight on memory,
 813         * we will try to reclaim all available objects, otherwise we can end
 814         * up failing allocations although there are plenty of reclaimable
 815         * objects spread over several slabs with usage less than the
 816         * batch_size.
 817         *
 818         * We detect the "tight on memory" situations by looking at the total
 819         * number of objects we want to scan (total_scan). If it is greater
 820         * than the total number of objects on slab (freeable), we must be
 821         * scanning at high prio and therefore should try to reclaim as much as
 822         * possible.
 823         */
 824        while (total_scan >= batch_size ||
 825               total_scan >= freeable) {
 826                unsigned long ret;
 827                unsigned long nr_to_scan = min(batch_size, total_scan);
 828
 829                shrinkctl->nr_to_scan = nr_to_scan;
 830                shrinkctl->nr_scanned = nr_to_scan;
 831                ret = shrinker->scan_objects(shrinker, shrinkctl);
 832                if (ret == SHRINK_STOP)
 833                        break;
 834                freed += ret;
 835
 836                count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
 837                total_scan -= shrinkctl->nr_scanned;
 838                scanned += shrinkctl->nr_scanned;
 839
 840                cond_resched();
 841        }
 842
 843        /*
 844         * The deferred work is increased by any new work (delta) that wasn't
 845         * done, decreased by old deferred work that was done now.
 846         *
 847         * And it is capped to two times of the freeable items.
 848         */
 849        next_deferred = max_t(long, (nr + delta - scanned), 0);
 850        next_deferred = min(next_deferred, (2 * freeable));
 851
 852        /*
 853         * move the unused scan count back into the shrinker in a
 854         * manner that handles concurrent updates.
 855         */
 856        new_nr = add_nr_deferred(next_deferred, shrinker, shrinkctl);
 857
 858        trace_mm_shrink_slab_end(shrinker, shrinkctl->nid, freed, nr, new_nr, total_scan);
 859        return freed;
 860}
 861
 862#ifdef CONFIG_MEMCG
 863static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
 864                        struct mem_cgroup *memcg, int priority)
 865{
 866        struct shrinker_info *info;
 867        unsigned long ret, freed = 0;
 868        int i;
 869
 870        if (!mem_cgroup_online(memcg))
 871                return 0;
 872
 873        if (!down_read_trylock(&shrinker_rwsem))
 874                return 0;
 875
 876        info = shrinker_info_protected(memcg, nid);
 877        if (unlikely(!info))
 878                goto unlock;
 879
 880        for_each_set_bit(i, info->map, shrinker_nr_max) {
 881                struct shrink_control sc = {
 882                        .gfp_mask = gfp_mask,
 883                        .nid = nid,
 884                        .memcg = memcg,
 885                };
 886                struct shrinker *shrinker;
 887
 888                shrinker = idr_find(&shrinker_idr, i);
 889                if (unlikely(!shrinker || !(shrinker->flags & SHRINKER_REGISTERED))) {
 890                        if (!shrinker)
 891                                clear_bit(i, info->map);
 892                        continue;
 893                }
 894
 895                /* Call non-slab shrinkers even though kmem is disabled */
 896                if (!memcg_kmem_enabled() &&
 897                    !(shrinker->flags & SHRINKER_NONSLAB))
 898                        continue;
 899
 900                ret = do_shrink_slab(&sc, shrinker, priority);
 901                if (ret == SHRINK_EMPTY) {
 902                        clear_bit(i, info->map);
 903                        /*
 904                         * After the shrinker reported that it had no objects to
 905                         * free, but before we cleared the corresponding bit in
 906                         * the memcg shrinker map, a new object might have been
 907                         * added. To make sure, we have the bit set in this
 908                         * case, we invoke the shrinker one more time and reset
 909                         * the bit if it reports that it is not empty anymore.
 910                         * The memory barrier here pairs with the barrier in
 911                         * set_shrinker_bit():
 912                         *
 913                         * list_lru_add()     shrink_slab_memcg()
 914                         *   list_add_tail()    clear_bit()
 915                         *   <MB>               <MB>
 916                         *   set_bit()          do_shrink_slab()
 917                         */
 918                        smp_mb__after_atomic();
 919                        ret = do_shrink_slab(&sc, shrinker, priority);
 920                        if (ret == SHRINK_EMPTY)
 921                                ret = 0;
 922                        else
 923                                set_shrinker_bit(memcg, nid, i);
 924                }
 925                freed += ret;
 926
 927                if (rwsem_is_contended(&shrinker_rwsem)) {
 928                        freed = freed ? : 1;
 929                        break;
 930                }
 931        }
 932unlock:
 933        up_read(&shrinker_rwsem);
 934        return freed;
 935}
 936#else /* CONFIG_MEMCG */
 937static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
 938                        struct mem_cgroup *memcg, int priority)
 939{
 940        return 0;
 941}
 942#endif /* CONFIG_MEMCG */
 943
 944/**
 945 * shrink_slab - shrink slab caches
 946 * @gfp_mask: allocation context
 947 * @nid: node whose slab caches to target
 948 * @memcg: memory cgroup whose slab caches to target
 949 * @priority: the reclaim priority
 950 *
 951 * Call the shrink functions to age shrinkable caches.
 952 *
 953 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
 954 * unaware shrinkers will receive a node id of 0 instead.
 955 *
 956 * @memcg specifies the memory cgroup to target. Unaware shrinkers
 957 * are called only if it is the root cgroup.
 958 *
 959 * @priority is sc->priority, we take the number of objects and >> by priority
 960 * in order to get the scan target.
 961 *
 962 * Returns the number of reclaimed slab objects.
 963 */
 964static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
 965                                 struct mem_cgroup *memcg,
 966                                 int priority)
 967{
 968        unsigned long ret, freed = 0;
 969        struct shrinker *shrinker;
 970
 971        /*
 972         * The root memcg might be allocated even though memcg is disabled
 973         * via "cgroup_disable=memory" boot parameter.  This could make
 974         * mem_cgroup_is_root() return false, then just run memcg slab
 975         * shrink, but skip global shrink.  This may result in premature
 976         * oom.
 977         */
 978        if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
 979                return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
 980
 981        if (!down_read_trylock(&shrinker_rwsem))
 982                goto out;
 983
 984        list_for_each_entry(shrinker, &shrinker_list, list) {
 985                struct shrink_control sc = {
 986                        .gfp_mask = gfp_mask,
 987                        .nid = nid,
 988                        .memcg = memcg,
 989                };
 990
 991                ret = do_shrink_slab(&sc, shrinker, priority);
 992                if (ret == SHRINK_EMPTY)
 993                        ret = 0;
 994                freed += ret;
 995                /*
 996                 * Bail out if someone want to register a new shrinker to
 997                 * prevent the registration from being stalled for long periods
 998                 * by parallel ongoing shrinking.
 999                 */
1000                if (rwsem_is_contended(&shrinker_rwsem)) {
1001                        freed = freed ? : 1;
1002                        break;
1003                }
1004        }
1005
1006        up_read(&shrinker_rwsem);
1007out:
1008        cond_resched();
1009        return freed;
1010}
1011
1012static void drop_slab_node(int nid)
1013{
1014        unsigned long freed;
1015        int shift = 0;
1016
1017        do {
1018                struct mem_cgroup *memcg = NULL;
1019
1020                if (fatal_signal_pending(current))
1021                        return;
1022
1023                freed = 0;
1024                memcg = mem_cgroup_iter(NULL, NULL, NULL);
1025                do {
1026                        freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
1027                } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
1028        } while ((freed >> shift++) > 1);
1029}
1030
1031void drop_slab(void)
1032{
1033        int nid;
1034
1035        for_each_online_node(nid)
1036                drop_slab_node(nid);
1037}
1038
1039static inline int is_page_cache_freeable(struct folio *folio)
1040{
1041        /*
1042         * A freeable page cache page is referenced only by the caller
1043         * that isolated the page, the page cache and optional buffer
1044         * heads at page->private.
1045         */
1046        return folio_ref_count(folio) - folio_test_private(folio) ==
1047                1 + folio_nr_pages(folio);
1048}
1049
1050/*
1051 * We detected a synchronous write error writing a folio out.  Probably
1052 * -ENOSPC.  We need to propagate that into the address_space for a subsequent
1053 * fsync(), msync() or close().
1054 *
1055 * The tricky part is that after writepage we cannot touch the mapping: nothing
1056 * prevents it from being freed up.  But we have a ref on the folio and once
1057 * that folio is locked, the mapping is pinned.
1058 *
1059 * We're allowed to run sleeping folio_lock() here because we know the caller has
1060 * __GFP_FS.
1061 */
1062static void handle_write_error(struct address_space *mapping,
1063                                struct folio *folio, int error)
1064{
1065        folio_lock(folio);
1066        if (folio_mapping(folio) == mapping)
1067                mapping_set_error(mapping, error);
1068        folio_unlock(folio);
1069}
1070
1071static bool skip_throttle_noprogress(pg_data_t *pgdat)
1072{
1073        int reclaimable = 0, write_pending = 0;
1074        int i;
1075
1076        /*
1077         * If kswapd is disabled, reschedule if necessary but do not
1078         * throttle as the system is likely near OOM.
1079         */
1080        if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
1081                return true;
1082
1083        /*
1084         * If there are a lot of dirty/writeback pages then do not
1085         * throttle as throttling will occur when the pages cycle
1086         * towards the end of the LRU if still under writeback.
1087         */
1088        for (i = 0; i < MAX_NR_ZONES; i++) {
1089                struct zone *zone = pgdat->node_zones + i;
1090
1091                if (!managed_zone(zone))
1092                        continue;
1093
1094                reclaimable += zone_reclaimable_pages(zone);
1095                write_pending += zone_page_state_snapshot(zone,
1096                                                  NR_ZONE_WRITE_PENDING);
1097        }
1098        if (2 * write_pending <= reclaimable)
1099                return true;
1100
1101        return false;
1102}
1103
1104void reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason)
1105{
1106        wait_queue_head_t *wqh = &pgdat->reclaim_wait[reason];
1107        long timeout, ret;
1108        DEFINE_WAIT(wait);
1109
1110        /*
1111         * Do not throttle IO workers, kthreads other than kswapd or
1112         * workqueues. They may be required for reclaim to make
1113         * forward progress (e.g. journalling workqueues or kthreads).
1114         */
1115        if (!current_is_kswapd() &&
1116            current->flags & (PF_IO_WORKER|PF_KTHREAD)) {
1117                cond_resched();
1118                return;
1119        }
1120
1121        /*
1122         * These figures are pulled out of thin air.
1123         * VMSCAN_THROTTLE_ISOLATED is a transient condition based on too many
1124         * parallel reclaimers which is a short-lived event so the timeout is
1125         * short. Failing to make progress or waiting on writeback are
1126         * potentially long-lived events so use a longer timeout. This is shaky
1127         * logic as a failure to make progress could be due to anything from
1128         * writeback to a slow device to excessive references pages at the tail
1129         * of the inactive LRU.
1130         */
1131        switch(reason) {
1132        case VMSCAN_THROTTLE_WRITEBACK:
1133                timeout = HZ/10;
1134
1135                if (atomic_inc_return(&pgdat->nr_writeback_throttled) == 1) {
1136                        WRITE_ONCE(pgdat->nr_reclaim_start,
1137                                node_page_state(pgdat, NR_THROTTLED_WRITTEN));
1138                }
1139
1140                break;
1141        case VMSCAN_THROTTLE_CONGESTED:
1142                fallthrough;
1143        case VMSCAN_THROTTLE_NOPROGRESS:
1144                if (skip_throttle_noprogress(pgdat)) {
1145                        cond_resched();
1146                        return;
1147                }
1148
1149                timeout = 1;
1150
1151                break;
1152        case VMSCAN_THROTTLE_ISOLATED:
1153                timeout = HZ/50;
1154                break;
1155        default:
1156                WARN_ON_ONCE(1);
1157                timeout = HZ;
1158                break;
1159        }
1160
1161        prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1162        ret = schedule_timeout(timeout);
1163        finish_wait(wqh, &wait);
1164
1165        if (reason == VMSCAN_THROTTLE_WRITEBACK)
1166                atomic_dec(&pgdat->nr_writeback_throttled);
1167
1168        trace_mm_vmscan_throttled(pgdat->node_id, jiffies_to_usecs(timeout),
1169                                jiffies_to_usecs(timeout - ret),
1170                                reason);
1171}
1172
1173/*
1174 * Account for pages written if tasks are throttled waiting on dirty
1175 * pages to clean. If enough pages have been cleaned since throttling
1176 * started then wakeup the throttled tasks.
1177 */
1178void __acct_reclaim_writeback(pg_data_t *pgdat, struct folio *folio,
1179                                                        int nr_throttled)
1180{
1181        unsigned long nr_written;
1182
1183        node_stat_add_folio(folio, NR_THROTTLED_WRITTEN);
1184
1185        /*
1186         * This is an inaccurate read as the per-cpu deltas may not
1187         * be synchronised. However, given that the system is
1188         * writeback throttled, it is not worth taking the penalty
1189         * of getting an accurate count. At worst, the throttle
1190         * timeout guarantees forward progress.
1191         */
1192        nr_written = node_page_state(pgdat, NR_THROTTLED_WRITTEN) -
1193                READ_ONCE(pgdat->nr_reclaim_start);
1194
1195        if (nr_written > SWAP_CLUSTER_MAX * nr_throttled)
1196                wake_up(&pgdat->reclaim_wait[VMSCAN_THROTTLE_WRITEBACK]);
1197}
1198
1199/* possible outcome of pageout() */
1200typedef enum {
1201        /* failed to write page out, page is locked */
1202        PAGE_KEEP,
1203        /* move page to the active list, page is locked */
1204        PAGE_ACTIVATE,
1205        /* page has been sent to the disk successfully, page is unlocked */
1206        PAGE_SUCCESS,
1207        /* page is clean and locked */
1208        PAGE_CLEAN,
1209} pageout_t;
1210
1211/*
1212 * pageout is called by shrink_page_list() for each dirty page.
1213 * Calls ->writepage().
1214 */
1215static pageout_t pageout(struct folio *folio, struct address_space *mapping,
1216                         struct swap_iocb **plug)
1217{
1218        /*
1219         * If the folio is dirty, only perform writeback if that write
1220         * will be non-blocking.  To prevent this allocation from being
1221         * stalled by pagecache activity.  But note that there may be
1222         * stalls if we need to run get_block().  We could test
1223         * PagePrivate for that.
1224         *
1225         * If this process is currently in __generic_file_write_iter() against
1226         * this folio's queue, we can perform writeback even if that
1227         * will block.
1228         *
1229         * If the folio is swapcache, write it back even if that would
1230         * block, for some throttling. This happens by accident, because
1231         * swap_backing_dev_info is bust: it doesn't reflect the
1232         * congestion state of the swapdevs.  Easy to fix, if needed.
1233         */
1234        if (!is_page_cache_freeable(folio))
1235                return PAGE_KEEP;
1236        if (!mapping) {
1237                /*
1238                 * Some data journaling orphaned folios can have
1239                 * folio->mapping == NULL while being dirty with clean buffers.
1240                 */
1241                if (folio_test_private(folio)) {
1242                        if (try_to_free_buffers(folio)) {
1243                                folio_clear_dirty(folio);
1244                                pr_info("%s: orphaned folio\n", __func__);
1245                                return PAGE_CLEAN;
1246                        }
1247                }
1248                return PAGE_KEEP;
1249        }
1250        if (mapping->a_ops->writepage == NULL)
1251                return PAGE_ACTIVATE;
1252
1253        if (folio_clear_dirty_for_io(folio)) {
1254                int res;
1255                struct writeback_control wbc = {
1256                        .sync_mode = WB_SYNC_NONE,
1257                        .nr_to_write = SWAP_CLUSTER_MAX,
1258                        .range_start = 0,
1259                        .range_end = LLONG_MAX,
1260                        .for_reclaim = 1,
1261                        .swap_plug = plug,
1262                };
1263
1264                folio_set_reclaim(folio);
1265                res = mapping->a_ops->writepage(&folio->page, &wbc);
1266                if (res < 0)
1267                        handle_write_error(mapping, folio, res);
1268                if (res == AOP_WRITEPAGE_ACTIVATE) {
1269                        folio_clear_reclaim(folio);
1270                        return PAGE_ACTIVATE;
1271                }
1272
1273                if (!folio_test_writeback(folio)) {
1274                        /* synchronous write or broken a_ops? */
1275                        folio_clear_reclaim(folio);
1276                }
1277                trace_mm_vmscan_write_folio(folio);
1278                node_stat_add_folio(folio, NR_VMSCAN_WRITE);
1279                return PAGE_SUCCESS;
1280        }
1281
1282        return PAGE_CLEAN;
1283}
1284
1285/*
1286 * Same as remove_mapping, but if the page is removed from the mapping, it
1287 * gets returned with a refcount of 0.
1288 */
1289static int __remove_mapping(struct address_space *mapping, struct folio *folio,
1290                            bool reclaimed, struct mem_cgroup *target_memcg)
1291{
1292        int refcount;
1293        void *shadow = NULL;
1294
1295        BUG_ON(!folio_test_locked(folio));
1296        BUG_ON(mapping != folio_mapping(folio));
1297
1298        if (!folio_test_swapcache(folio))
1299                spin_lock(&mapping->host->i_lock);
1300        xa_lock_irq(&mapping->i_pages);
1301        /*
1302         * The non racy check for a busy page.
1303         *
1304         * Must be careful with the order of the tests. When someone has
1305         * a ref to the page, it may be possible that they dirty it then
1306         * drop the reference. So if PageDirty is tested before page_count
1307         * here, then the following race may occur:
1308         *
1309         * get_user_pages(&page);
1310         * [user mapping goes away]
1311         * write_to(page);
1312         *                              !PageDirty(page)    [good]
1313         * SetPageDirty(page);
1314         * put_page(page);
1315         *                              !page_count(page)   [good, discard it]
1316         *
1317         * [oops, our write_to data is lost]
1318         *
1319         * Reversing the order of the tests ensures such a situation cannot
1320         * escape unnoticed. The smp_rmb is needed to ensure the page->flags
1321         * load is not satisfied before that of page->_refcount.
1322         *
1323         * Note that if SetPageDirty is always performed via set_page_dirty,
1324         * and thus under the i_pages lock, then this ordering is not required.
1325         */
1326        refcount = 1 + folio_nr_pages(folio);
1327        if (!folio_ref_freeze(folio, refcount))
1328                goto cannot_free;
1329        /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
1330        if (unlikely(folio_test_dirty(folio))) {
1331                folio_ref_unfreeze(folio, refcount);
1332                goto cannot_free;
1333        }
1334
1335        if (folio_test_swapcache(folio)) {
1336                swp_entry_t swap = folio_swap_entry(folio);
1337                mem_cgroup_swapout(folio, swap);
1338                if (reclaimed && !mapping_exiting(mapping))
1339                        shadow = workingset_eviction(folio, target_memcg);
1340                __delete_from_swap_cache(folio, swap, shadow);
1341                xa_unlock_irq(&mapping->i_pages);
1342                put_swap_page(&folio->page, swap);
1343        } else {
1344                void (*free_folio)(struct folio *);
1345
1346                free_folio = mapping->a_ops->free_folio;
1347                /*
1348                 * Remember a shadow entry for reclaimed file cache in
1349                 * order to detect refaults, thus thrashing, later on.
1350                 *
1351                 * But don't store shadows in an address space that is
1352                 * already exiting.  This is not just an optimization,
1353                 * inode reclaim needs to empty out the radix tree or
1354                 * the nodes are lost.  Don't plant shadows behind its
1355                 * back.
1356                 *
1357                 * We also don't store shadows for DAX mappings because the
1358                 * only page cache pages found in these are zero pages
1359                 * covering holes, and because we don't want to mix DAX
1360                 * exceptional entries and shadow exceptional entries in the
1361                 * same address_space.
1362                 */
1363                if (reclaimed && folio_is_file_lru(folio) &&
1364                    !mapping_exiting(mapping) && !dax_mapping(mapping))
1365                        shadow = workingset_eviction(folio, target_memcg);
1366                __filemap_remove_folio(folio, shadow);
1367                xa_unlock_irq(&mapping->i_pages);
1368                if (mapping_shrinkable(mapping))
1369                        inode_add_lru(mapping->host);
1370                spin_unlock(&mapping->host->i_lock);
1371
1372                if (free_folio)
1373                        free_folio(folio);
1374        }
1375
1376        return 1;
1377
1378cannot_free:
1379        xa_unlock_irq(&mapping->i_pages);
1380        if (!folio_test_swapcache(folio))
1381                spin_unlock(&mapping->host->i_lock);
1382        return 0;
1383}
1384
1385/**
1386 * remove_mapping() - Attempt to remove a folio from its mapping.
1387 * @mapping: The address space.
1388 * @folio: The folio to remove.
1389 *
1390 * If the folio is dirty, under writeback or if someone else has a ref
1391 * on it, removal will fail.
1392 * Return: The number of pages removed from the mapping.  0 if the folio
1393 * could not be removed.
1394 * Context: The caller should have a single refcount on the folio and
1395 * hold its lock.
1396 */
1397long remove_mapping(struct address_space *mapping, struct folio *folio)
1398{
1399        if (__remove_mapping(mapping, folio, false, NULL)) {
1400                /*
1401                 * Unfreezing the refcount with 1 effectively
1402                 * drops the pagecache ref for us without requiring another
1403                 * atomic operation.
1404                 */
1405                folio_ref_unfreeze(folio, 1);
1406                return folio_nr_pages(folio);
1407        }
1408        return 0;
1409}
1410
1411/**
1412 * folio_putback_lru - Put previously isolated folio onto appropriate LRU list.
1413 * @folio: Folio to be returned to an LRU list.
1414 *
1415 * Add previously isolated @folio to appropriate LRU list.
1416 * The folio may still be unevictable for other reasons.
1417 *
1418 * Context: lru_lock must not be held, interrupts must be enabled.
1419 */
1420void folio_putback_lru(struct folio *folio)
1421{
1422        folio_add_lru(folio);
1423        folio_put(folio);               /* drop ref from isolate */
1424}
1425
1426enum page_references {
1427        PAGEREF_RECLAIM,
1428        PAGEREF_RECLAIM_CLEAN,
1429        PAGEREF_KEEP,
1430        PAGEREF_ACTIVATE,
1431};
1432
1433static enum page_references folio_check_references(struct folio *folio,
1434                                                  struct scan_control *sc)
1435{
1436        int referenced_ptes, referenced_folio;
1437        unsigned long vm_flags;
1438
1439        referenced_ptes = folio_referenced(folio, 1, sc->target_mem_cgroup,
1440                                           &vm_flags);
1441        referenced_folio = folio_test_clear_referenced(folio);
1442
1443        /*
1444         * The supposedly reclaimable folio was found to be in a VM_LOCKED vma.
1445         * Let the folio, now marked Mlocked, be moved to the unevictable list.
1446         */
1447        if (vm_flags & VM_LOCKED)
1448                return PAGEREF_ACTIVATE;
1449
1450        /* rmap lock contention: rotate */
1451        if (referenced_ptes == -1)
1452                return PAGEREF_KEEP;
1453
1454        if (referenced_ptes) {
1455                /*
1456                 * All mapped folios start out with page table
1457                 * references from the instantiating fault, so we need
1458                 * to look twice if a mapped file/anon folio is used more
1459                 * than once.
1460                 *
1461                 * Mark it and spare it for another trip around the
1462                 * inactive list.  Another page table reference will
1463                 * lead to its activation.
1464                 *
1465                 * Note: the mark is set for activated folios as well
1466                 * so that recently deactivated but used folios are
1467                 * quickly recovered.
1468                 */
1469                folio_set_referenced(folio);
1470
1471                if (referenced_folio || referenced_ptes > 1)
1472                        return PAGEREF_ACTIVATE;
1473
1474                /*
1475                 * Activate file-backed executable folios after first usage.
1476                 */
1477                if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio))
1478                        return PAGEREF_ACTIVATE;
1479
1480                return PAGEREF_KEEP;
1481        }
1482
1483        /* Reclaim if clean, defer dirty folios to writeback */
1484        if (referenced_folio && folio_is_file_lru(folio))
1485                return PAGEREF_RECLAIM_CLEAN;
1486
1487        return PAGEREF_RECLAIM;
1488}
1489
1490/* Check if a page is dirty or under writeback */
1491static void folio_check_dirty_writeback(struct folio *folio,
1492                                       bool *dirty, bool *writeback)
1493{
1494        struct address_space *mapping;
1495
1496        /*
1497         * Anonymous pages are not handled by flushers and must be written
1498         * from reclaim context. Do not stall reclaim based on them.
1499         * MADV_FREE anonymous pages are put into inactive file list too.
1500         * They could be mistakenly treated as file lru. So further anon
1501         * test is needed.
1502         */
1503        if (!folio_is_file_lru(folio) ||
1504            (folio_test_anon(folio) && !folio_test_swapbacked(folio))) {
1505                *dirty = false;
1506                *writeback = false;
1507                return;
1508        }
1509
1510        /* By default assume that the folio flags are accurate */
1511        *dirty = folio_test_dirty(folio);
1512        *writeback = folio_test_writeback(folio);
1513
1514        /* Verify dirty/writeback state if the filesystem supports it */
1515        if (!folio_test_private(folio))
1516                return;
1517
1518        mapping = folio_mapping(folio);
1519        if (mapping && mapping->a_ops->is_dirty_writeback)
1520                mapping->a_ops->is_dirty_writeback(folio, dirty, writeback);
1521}
1522
1523static struct page *alloc_demote_page(struct page *page, unsigned long node)
1524{
1525        struct migration_target_control mtc = {
1526                /*
1527                 * Allocate from 'node', or fail quickly and quietly.
1528                 * When this happens, 'page' will likely just be discarded
1529                 * instead of migrated.
1530                 */
1531                .gfp_mask = (GFP_HIGHUSER_MOVABLE & ~__GFP_RECLAIM) |
1532                            __GFP_THISNODE  | __GFP_NOWARN |
1533                            __GFP_NOMEMALLOC | GFP_NOWAIT,
1534                .nid = node
1535        };
1536
1537        return alloc_migration_target(page, (unsigned long)&mtc);
1538}
1539
1540/*
1541 * Take pages on @demote_list and attempt to demote them to
1542 * another node.  Pages which are not demoted are left on
1543 * @demote_pages.
1544 */
1545static unsigned int demote_page_list(struct list_head *demote_pages,
1546                                     struct pglist_data *pgdat)
1547{
1548        int target_nid = next_demotion_node(pgdat->node_id);
1549        unsigned int nr_succeeded;
1550
1551        if (list_empty(demote_pages))
1552                return 0;
1553
1554        if (target_nid == NUMA_NO_NODE)
1555                return 0;
1556
1557        /* Demotion ignores all cpuset and mempolicy settings */
1558        migrate_pages(demote_pages, alloc_demote_page, NULL,
1559                            target_nid, MIGRATE_ASYNC, MR_DEMOTION,
1560                            &nr_succeeded);
1561
1562        if (current_is_kswapd())
1563                __count_vm_events(PGDEMOTE_KSWAPD, nr_succeeded);
1564        else
1565                __count_vm_events(PGDEMOTE_DIRECT, nr_succeeded);
1566
1567        return nr_succeeded;
1568}
1569
1570static bool may_enter_fs(struct folio *folio, gfp_t gfp_mask)
1571{
1572        if (gfp_mask & __GFP_FS)
1573                return true;
1574        if (!folio_test_swapcache(folio) || !(gfp_mask & __GFP_IO))
1575                return false;
1576        /*
1577         * We can "enter_fs" for swap-cache with only __GFP_IO
1578         * providing this isn't SWP_FS_OPS.
1579         * ->flags can be updated non-atomicially (scan_swap_map_slots),
1580         * but that will never affect SWP_FS_OPS, so the data_race
1581         * is safe.
1582         */
1583        return !data_race(folio_swap_flags(folio) & SWP_FS_OPS);
1584}
1585
1586/*
1587 * shrink_page_list() returns the number of reclaimed pages
1588 */
1589static unsigned int shrink_page_list(struct list_head *page_list,
1590                                     struct pglist_data *pgdat,
1591                                     struct scan_control *sc,
1592                                     struct reclaim_stat *stat,
1593                                     bool ignore_references)
1594{
1595        LIST_HEAD(ret_pages);
1596        LIST_HEAD(free_pages);
1597        LIST_HEAD(demote_pages);
1598        unsigned int nr_reclaimed = 0;
1599        unsigned int pgactivate = 0;
1600        bool do_demote_pass;
1601        struct swap_iocb *plug = NULL;
1602
1603        memset(stat, 0, sizeof(*stat));
1604        cond_resched();
1605        do_demote_pass = can_demote(pgdat->node_id, sc);
1606
1607retry:
1608        while (!list_empty(page_list)) {
1609                struct address_space *mapping;
1610                struct folio *folio;
1611                enum page_references references = PAGEREF_RECLAIM;
1612                bool dirty, writeback;
1613                unsigned int nr_pages;
1614
1615                cond_resched();
1616
1617                folio = lru_to_folio(page_list);
1618                list_del(&folio->lru);
1619
1620                if (!folio_trylock(folio))
1621                        goto keep;
1622
1623                VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
1624
1625                nr_pages = folio_nr_pages(folio);
1626
1627                /* Account the number of base pages */
1628                sc->nr_scanned += nr_pages;
1629
1630                if (unlikely(!folio_evictable(folio)))
1631                        goto activate_locked;
1632
1633                if (!sc->may_unmap && folio_mapped(folio))
1634                        goto keep_locked;
1635
1636                /*
1637                 * The number of dirty pages determines if a node is marked
1638                 * reclaim_congested. kswapd will stall and start writing
1639                 * folios if the tail of the LRU is all dirty unqueued folios.
1640                 */
1641                folio_check_dirty_writeback(folio, &dirty, &writeback);
1642                if (dirty || writeback)
1643                        stat->nr_dirty += nr_pages;
1644
1645                if (dirty && !writeback)
1646                        stat->nr_unqueued_dirty += nr_pages;
1647
1648                /*
1649                 * Treat this folio as congested if folios are cycling
1650                 * through the LRU so quickly that the folios marked
1651                 * for immediate reclaim are making it to the end of
1652                 * the LRU a second time.
1653                 */
1654                if (writeback && folio_test_reclaim(folio))
1655                        stat->nr_congested += nr_pages;
1656
1657                /*
1658                 * If a folio at the tail of the LRU is under writeback, there
1659                 * are three cases to consider.
1660                 *
1661                 * 1) If reclaim is encountering an excessive number
1662                 *    of folios under writeback and this folio has both
1663                 *    the writeback and reclaim flags set, then it
1664                 *    indicates that folios are being queued for I/O but
1665                 *    are being recycled through the LRU before the I/O
1666                 *    can complete. Waiting on the folio itself risks an
1667                 *    indefinite stall if it is impossible to writeback
1668                 *    the folio due to I/O error or disconnected storage
1669                 *    so instead note that the LRU is being scanned too
1670                 *    quickly and the caller can stall after the folio
1671                 *    list has been processed.
1672                 *
1673                 * 2) Global or new memcg reclaim encounters a folio that is
1674                 *    not marked for immediate reclaim, or the caller does not
1675                 *    have __GFP_FS (or __GFP_IO if it's simply going to swap,
1676                 *    not to fs). In this case mark the folio for immediate
1677                 *    reclaim and continue scanning.
1678                 *
1679                 *    Require may_enter_fs() because we would wait on fs, which
1680                 *    may not have submitted I/O yet. And the loop driver might
1681                 *    enter reclaim, and deadlock if it waits on a folio for
1682                 *    which it is needed to do the write (loop masks off
1683                 *    __GFP_IO|__GFP_FS for this reason); but more thought
1684                 *    would probably show more reasons.
1685                 *
1686                 * 3) Legacy memcg encounters a folio that already has the
1687                 *    reclaim flag set. memcg does not have any dirty folio
1688                 *    throttling so we could easily OOM just because too many
1689                 *    folios are in writeback and there is nothing else to
1690                 *    reclaim. Wait for the writeback to complete.
1691                 *
1692                 * In cases 1) and 2) we activate the folios to get them out of
1693                 * the way while we continue scanning for clean folios on the
1694                 * inactive list and refilling from the active list. The
1695                 * observation here is that waiting for disk writes is more
1696                 * expensive than potentially causing reloads down the line.
1697                 * Since they're marked for immediate reclaim, they won't put
1698                 * memory pressure on the cache working set any longer than it
1699                 * takes to write them to disk.
1700                 */
1701                if (folio_test_writeback(folio)) {
1702                        /* Case 1 above */
1703                        if (current_is_kswapd() &&
1704                            folio_test_reclaim(folio) &&
1705                            test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1706                                stat->nr_immediate += nr_pages;
1707                                goto activate_locked;
1708
1709                        /* Case 2 above */
1710                        } else if (writeback_throttling_sane(sc) ||
1711                            !folio_test_reclaim(folio) ||
1712                            !may_enter_fs(folio, sc->gfp_mask)) {
1713                                /*
1714                                 * This is slightly racy -
1715                                 * folio_end_writeback() might have
1716                                 * just cleared the reclaim flag, then
1717                                 * setting the reclaim flag here ends up
1718                                 * interpreted as the readahead flag - but
1719                                 * that does not matter enough to care.
1720                                 * What we do want is for this folio to
1721                                 * have the reclaim flag set next time
1722                                 * memcg reclaim reaches the tests above,
1723                                 * so it will then wait for writeback to
1724                                 * avoid OOM; and it's also appropriate
1725                                 * in global reclaim.
1726                                 */
1727                                folio_set_reclaim(folio);
1728                                stat->nr_writeback += nr_pages;
1729                                goto activate_locked;
1730
1731                        /* Case 3 above */
1732                        } else {
1733                                folio_unlock(folio);
1734                                folio_wait_writeback(folio);
1735                                /* then go back and try same folio again */
1736                                list_add_tail(&folio->lru, page_list);
1737                                continue;
1738                        }
1739                }
1740
1741                if (!ignore_references)
1742                        references = folio_check_references(folio, sc);
1743
1744                switch (references) {
1745                case PAGEREF_ACTIVATE:
1746                        goto activate_locked;
1747                case PAGEREF_KEEP:
1748                        stat->nr_ref_keep += nr_pages;
1749                        goto keep_locked;
1750                case PAGEREF_RECLAIM:
1751                case PAGEREF_RECLAIM_CLEAN:
1752                        ; /* try to reclaim the folio below */
1753                }
1754
1755                /*
1756                 * Before reclaiming the folio, try to relocate
1757                 * its contents to another node.
1758                 */
1759                if (do_demote_pass &&
1760                    (thp_migration_supported() || !folio_test_large(folio))) {
1761                        list_add(&folio->lru, &demote_pages);
1762                        folio_unlock(folio);
1763                        continue;
1764                }
1765
1766                /*
1767                 * Anonymous process memory has backing store?
1768                 * Try to allocate it some swap space here.
1769                 * Lazyfree folio could be freed directly
1770                 */
1771                if (folio_test_anon(folio) && folio_test_swapbacked(folio)) {
1772                        if (!folio_test_swapcache(folio)) {
1773                                if (!(sc->gfp_mask & __GFP_IO))
1774                                        goto keep_locked;
1775                                if (folio_maybe_dma_pinned(folio))
1776                                        goto keep_locked;
1777                                if (folio_test_large(folio)) {
1778                                        /* cannot split folio, skip it */
1779                                        if (!can_split_folio(folio, NULL))
1780                                                goto activate_locked;
1781                                        /*
1782                                         * Split folios without a PMD map right
1783                                         * away. Chances are some or all of the
1784                                         * tail pages can be freed without IO.
1785                                         */
1786                                        if (!folio_entire_mapcount(folio) &&
1787                                            split_folio_to_list(folio,
1788                                                                page_list))
1789                                                goto activate_locked;
1790                                }
1791                                if (!add_to_swap(folio)) {
1792                                        if (!folio_test_large(folio))
1793                                                goto activate_locked_split;
1794                                        /* Fallback to swap normal pages */
1795                                        if (split_folio_to_list(folio,
1796                                                                page_list))
1797                                                goto activate_locked;
1798#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1799                                        count_vm_event(THP_SWPOUT_FALLBACK);
1800#endif
1801                                        if (!add_to_swap(folio))
1802                                                goto activate_locked_split;
1803                                }
1804                        }
1805                } else if (folio_test_swapbacked(folio) &&
1806                           folio_test_large(folio)) {
1807                        /* Split shmem folio */
1808                        if (split_folio_to_list(folio, page_list))
1809                                goto keep_locked;
1810                }
1811
1812                /*
1813                 * If the folio was split above, the tail pages will make
1814                 * their own pass through this function and be accounted
1815                 * then.
1816                 */
1817                if ((nr_pages > 1) && !folio_test_large(folio)) {
1818                        sc->nr_scanned -= (nr_pages - 1);
1819                        nr_pages = 1;
1820                }
1821
1822                /*
1823                 * The folio is mapped into the page tables of one or more
1824                 * processes. Try to unmap it here.
1825                 */
1826                if (folio_mapped(folio)) {
1827                        enum ttu_flags flags = TTU_BATCH_FLUSH;
1828                        bool was_swapbacked = folio_test_swapbacked(folio);
1829
1830                        if (folio_test_pmd_mappable(folio))
1831                                flags |= TTU_SPLIT_HUGE_PMD;
1832
1833                        try_to_unmap(folio, flags);
1834                        if (folio_mapped(folio)) {
1835                                stat->nr_unmap_fail += nr_pages;
1836                                if (!was_swapbacked &&
1837                                    folio_test_swapbacked(folio))
1838                                        stat->nr_lazyfree_fail += nr_pages;
1839                                goto activate_locked;
1840                        }
1841                }
1842
1843                mapping = folio_mapping(folio);
1844                if (folio_test_dirty(folio)) {
1845                        /*
1846                         * Only kswapd can writeback filesystem folios
1847                         * to avoid risk of stack overflow. But avoid
1848                         * injecting inefficient single-folio I/O into
1849                         * flusher writeback as much as possible: only
1850                         * write folios when we've encountered many
1851                         * dirty folios, and when we've already scanned
1852                         * the rest of the LRU for clean folios and see
1853                         * the same dirty folios again (with the reclaim
1854                         * flag set).
1855                         */
1856                        if (folio_is_file_lru(folio) &&
1857                            (!current_is_kswapd() ||
1858                             !folio_test_reclaim(folio) ||
1859                             !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1860                                /*
1861                                 * Immediately reclaim when written back.
1862                                 * Similar in principle to deactivate_page()
1863                                 * except we already have the folio isolated
1864                                 * and know it's dirty
1865                                 */
1866                                node_stat_mod_folio(folio, NR_VMSCAN_IMMEDIATE,
1867                                                nr_pages);
1868                                folio_set_reclaim(folio);
1869
1870                                goto activate_locked;
1871                        }
1872
1873                        if (references == PAGEREF_RECLAIM_CLEAN)
1874                                goto keep_locked;
1875                        if (!may_enter_fs(folio, sc->gfp_mask))
1876                                goto keep_locked;
1877                        if (!sc->may_writepage)
1878                                goto keep_locked;
1879
1880                        /*
1881                         * Folio is dirty. Flush the TLB if a writable entry
1882                         * potentially exists to avoid CPU writes after I/O
1883                         * starts and then write it out here.
1884                         */
1885                        try_to_unmap_flush_dirty();
1886                        switch (pageout(folio, mapping, &plug)) {
1887                        case PAGE_KEEP:
1888                                goto keep_locked;
1889                        case PAGE_ACTIVATE:
1890                                goto activate_locked;
1891                        case PAGE_SUCCESS:
1892                                stat->nr_pageout += nr_pages;
1893
1894                                if (folio_test_writeback(folio))
1895                                        goto keep;
1896                                if (folio_test_dirty(folio))
1897                                        goto keep;
1898
1899                                /*
1900                                 * A synchronous write - probably a ramdisk.  Go
1901                                 * ahead and try to reclaim the folio.
1902                                 */
1903                                if (!folio_trylock(folio))
1904                                        goto keep;
1905                                if (folio_test_dirty(folio) ||
1906                                    folio_test_writeback(folio))
1907                                        goto keep_locked;
1908                                mapping = folio_mapping(folio);
1909                                fallthrough;
1910                        case PAGE_CLEAN:
1911                                ; /* try to free the folio below */
1912                        }
1913                }
1914
1915                /*
1916                 * If the folio has buffers, try to free the buffer
1917                 * mappings associated with this folio. If we succeed
1918                 * we try to free the folio as well.
1919                 *
1920                 * We do this even if the folio is dirty.
1921                 * filemap_release_folio() does not perform I/O, but it
1922                 * is possible for a folio to have the dirty flag set,
1923                 * but it is actually clean (all its buffers are clean).
1924                 * This happens if the buffers were written out directly,
1925                 * with submit_bh(). ext3 will do this, as well as
1926                 * the blockdev mapping.  filemap_release_folio() will
1927                 * discover that cleanness and will drop the buffers
1928                 * and mark the folio clean - it can be freed.
1929                 *
1930                 * Rarely, folios can have buffers and no ->mapping.
1931                 * These are the folios which were not successfully
1932                 * invalidated in truncate_cleanup_folio().  We try to
1933                 * drop those buffers here and if that worked, and the
1934                 * folio is no longer mapped into process address space
1935                 * (refcount == 1) it can be freed.  Otherwise, leave
1936                 * the folio on the LRU so it is swappable.
1937                 */
1938                if (folio_has_private(folio)) {
1939                        if (!filemap_release_folio(folio, sc->gfp_mask))
1940                                goto activate_locked;
1941                        if (!mapping && folio_ref_count(folio) == 1) {
1942                                folio_unlock(folio);
1943                                if (folio_put_testzero(folio))
1944                                        goto free_it;
1945                                else {
1946                                        /*
1947                                         * rare race with speculative reference.
1948                                         * the speculative reference will free
1949                                         * this folio shortly, so we may
1950                                         * increment nr_reclaimed here (and
1951                                         * leave it off the LRU).
1952                                         */
1953                                        nr_reclaimed += nr_pages;
1954                                        continue;
1955                                }
1956                        }
1957                }
1958
1959                if (folio_test_anon(folio) && !folio_test_swapbacked(folio)) {
1960                        /* follow __remove_mapping for reference */
1961                        if (!folio_ref_freeze(folio, 1))
1962                                goto keep_locked;
1963                        /*
1964                         * The folio has only one reference left, which is
1965                         * from the isolation. After the caller puts the
1966                         * folio back on the lru and drops the reference, the
1967                         * folio will be freed anyway. It doesn't matter
1968                         * which lru it goes on. So we don't bother checking
1969                         * the dirty flag here.
1970                         */
1971                        count_vm_events(PGLAZYFREED, nr_pages);
1972                        count_memcg_folio_events(folio, PGLAZYFREED, nr_pages);
1973                } else if (!mapping || !__remove_mapping(mapping, folio, true,
1974                                                         sc->target_mem_cgroup))
1975                        goto keep_locked;
1976
1977                folio_unlock(folio);
1978free_it:
1979                /*
1980                 * Folio may get swapped out as a whole, need to account
1981                 * all pages in it.
1982                 */
1983                nr_reclaimed += nr_pages;
1984
1985                /*
1986                 * Is there need to periodically free_page_list? It would
1987                 * appear not as the counts should be low
1988                 */
1989                if (unlikely(folio_test_large(folio)))
1990                        destroy_large_folio(folio);
1991                else
1992                        list_add(&folio->lru, &free_pages);
1993                continue;
1994
1995activate_locked_split:
1996                /*
1997                 * The tail pages that are failed to add into swap cache
1998                 * reach here.  Fixup nr_scanned and nr_pages.
1999                 */
2000                if (nr_pages > 1) {
2001                        sc->nr_scanned -= (nr_pages - 1);
2002                        nr_pages = 1;
2003                }
2004activate_locked:
2005                /* Not a candidate for swapping, so reclaim swap space. */
2006                if (folio_test_swapcache(folio) &&
2007                    (mem_cgroup_swap_full(&folio->page) ||
2008                     folio_test_mlocked(folio)))
2009                        try_to_free_swap(&folio->page);
2010                VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
2011                if (!folio_test_mlocked(folio)) {
2012                        int type = folio_is_file_lru(folio);
2013                        folio_set_active(folio);
2014                        stat->nr_activate[type] += nr_pages;
2015                        count_memcg_folio_events(folio, PGACTIVATE, nr_pages);
2016                }
2017keep_locked:
2018                folio_unlock(folio);
2019keep:
2020                list_add(&folio->lru, &ret_pages);
2021                VM_BUG_ON_FOLIO(folio_test_lru(folio) ||
2022                                folio_test_unevictable(folio), folio);
2023        }
2024        /* 'page_list' is always empty here */
2025
2026        /* Migrate folios selected for demotion */
2027        nr_reclaimed += demote_page_list(&demote_pages, pgdat);
2028        /* Folios that could not be demoted are still in @demote_pages */
2029        if (!list_empty(&demote_pages)) {
2030                /* Folios which weren't demoted go back on @page_list for retry: */
2031                list_splice_init(&demote_pages, page_list);
2032                do_demote_pass = false;
2033                goto retry;
2034        }
2035
2036        pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
2037
2038        mem_cgroup_uncharge_list(&free_pages);
2039        try_to_unmap_flush();
2040        free_unref_page_list(&free_pages);
2041
2042        list_splice(&ret_pages, page_list);
2043        count_vm_events(PGACTIVATE, pgactivate);
2044
2045        if (plug)
2046                swap_write_unplug(plug);
2047        return nr_reclaimed;
2048}
2049
2050unsigned int reclaim_clean_pages_from_list(struct zone *zone,
2051                                            struct list_head *folio_list)
2052{
2053        struct scan_control sc = {
2054                .gfp_mask = GFP_KERNEL,
2055                .may_unmap = 1,
2056        };
2057        struct reclaim_stat stat;
2058        unsigned int nr_reclaimed;
2059        struct folio *folio, *next;
2060        LIST_HEAD(clean_folios);
2061        unsigned int noreclaim_flag;
2062
2063        list_for_each_entry_safe(folio, next, folio_list, lru) {
2064                if (!folio_test_hugetlb(folio) && folio_is_file_lru(folio) &&
2065                    !folio_test_dirty(folio) && !__folio_test_movable(folio) &&
2066                    !folio_test_unevictable(folio)) {
2067                        folio_clear_active(folio);
2068                        list_move(&folio->lru, &clean_folios);
2069                }
2070        }
2071
2072        /*
2073         * We should be safe here since we are only dealing with file pages and
2074         * we are not kswapd and therefore cannot write dirty file pages. But
2075         * call memalloc_noreclaim_save() anyway, just in case these conditions
2076         * change in the future.
2077         */
2078        noreclaim_flag = memalloc_noreclaim_save();
2079        nr_reclaimed = shrink_page_list(&clean_folios, zone->zone_pgdat, &sc,
2080                                        &stat, true);
2081        memalloc_noreclaim_restore(noreclaim_flag);
2082
2083        list_splice(&clean_folios, folio_list);
2084        mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
2085                            -(long)nr_reclaimed);
2086        /*
2087         * Since lazyfree pages are isolated from file LRU from the beginning,
2088         * they will rotate back to anonymous LRU in the end if it failed to
2089         * discard so isolated count will be mismatched.
2090         * Compensate the isolated count for both LRU lists.
2091         */
2092        mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
2093                            stat.nr_lazyfree_fail);
2094        mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
2095                            -(long)stat.nr_lazyfree_fail);
2096        return nr_reclaimed;
2097}
2098
2099/*
2100 * Update LRU sizes after isolating pages. The LRU size updates must
2101 * be complete before mem_cgroup_update_lru_size due to a sanity check.
2102 */
2103static __always_inline void update_lru_sizes(struct lruvec *lruvec,
2104                        enum lru_list lru, unsigned long *nr_zone_taken)
2105{
2106        int zid;
2107
2108        for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2109                if (!nr_zone_taken[zid])
2110                        continue;
2111
2112                update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
2113        }
2114
2115}
2116
2117/*
2118 * Isolating page from the lruvec to fill in @dst list by nr_to_scan times.
2119 *
2120 * lruvec->lru_lock is heavily contended.  Some of the functions that
2121 * shrink the lists perform better by taking out a batch of pages
2122 * and working on them outside the LRU lock.
2123 *
2124 * For pagecache intensive workloads, this function is the hottest
2125 * spot in the kernel (apart from copy_*_user functions).
2126 *
2127 * Lru_lock must be held before calling this function.
2128 *
2129 * @nr_to_scan: The number of eligible pages to look through on the list.
2130 * @lruvec:     The LRU vector to pull pages from.
2131 * @dst:        The temp list to put pages on to.
2132 * @nr_scanned: The number of pages that were scanned.
2133 * @sc:         The scan_control struct for this reclaim session
2134 * @lru:        LRU list id for isolating
2135 *
2136 * returns how many pages were moved onto *@dst.
2137 */
2138static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
2139                struct lruvec *lruvec, struct list_head *dst,
2140                unsigned long *nr_scanned, struct scan_control *sc,
2141                enum lru_list lru)
2142{
2143        struct list_head *src = &lruvec->lists[lru];
2144        unsigned long nr_taken = 0;
2145        unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
2146        unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
2147        unsigned long skipped = 0;
2148        unsigned long scan, total_scan, nr_pages;
2149        LIST_HEAD(folios_skipped);
2150
2151        total_scan = 0;
2152        scan = 0;
2153        while (scan < nr_to_scan && !list_empty(src)) {
2154                struct list_head *move_to = src;
2155                struct folio *folio;
2156
2157                folio = lru_to_folio(src);
2158                prefetchw_prev_lru_folio(folio, src, flags);
2159
2160                nr_pages = folio_nr_pages(folio);
2161                total_scan += nr_pages;
2162
2163                if (folio_zonenum(folio) > sc->reclaim_idx) {
2164                        nr_skipped[folio_zonenum(folio)] += nr_pages;
2165                        move_to = &folios_skipped;
2166                        goto move;
2167                }
2168
2169                /*
2170                 * Do not count skipped folios because that makes the function
2171                 * return with no isolated folios if the LRU mostly contains
2172                 * ineligible folios.  This causes the VM to not reclaim any
2173                 * folios, triggering a premature OOM.
2174                 * Account all pages in a folio.
2175                 */
2176                scan += nr_pages;
2177
2178                if (!folio_test_lru(folio))
2179                        goto move;
2180                if (!sc->may_unmap && folio_mapped(folio))
2181                        goto move;
2182
2183                /*
2184                 * Be careful not to clear the lru flag until after we're
2185                 * sure the folio is not being freed elsewhere -- the
2186                 * folio release code relies on it.
2187                 */
2188                if (unlikely(!folio_try_get(folio)))
2189                        goto move;
2190
2191                if (!folio_test_clear_lru(folio)) {
2192                        /* Another thread is already isolating this folio */
2193                        folio_put(folio);
2194                        goto move;
2195                }
2196
2197                nr_taken += nr_pages;
2198                nr_zone_taken[folio_zonenum(folio)] += nr_pages;
2199                move_to = dst;
2200move:
2201                list_move(&folio->lru, move_to);
2202        }
2203
2204        /*
2205         * Splice any skipped folios to the start of the LRU list. Note that
2206         * this disrupts the LRU order when reclaiming for lower zones but
2207         * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
2208         * scanning would soon rescan the same folios to skip and waste lots
2209         * of cpu cycles.
2210         */
2211        if (!list_empty(&folios_skipped)) {
2212                int zid;
2213
2214                list_splice(&folios_skipped, src);
2215                for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2216                        if (!nr_skipped[zid])
2217                                continue;
2218
2219                        __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
2220                        skipped += nr_skipped[zid];
2221                }
2222        }
2223        *nr_scanned = total_scan;
2224        trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
2225                                    total_scan, skipped, nr_taken,
2226                                    sc->may_unmap ? 0 : ISOLATE_UNMAPPED, lru);
2227        update_lru_sizes(lruvec, lru, nr_zone_taken);
2228        return nr_taken;
2229}
2230
2231/**
2232 * folio_isolate_lru() - Try to isolate a folio from its LRU list.
2233 * @folio: Folio to isolate from its LRU list.
2234 *
2235 * Isolate a @folio from an LRU list and adjust the vmstat statistic
2236 * corresponding to whatever LRU list the folio was on.
2237 *
2238 * The folio will have its LRU flag cleared.  If it was found on the
2239 * active list, it will have the Active flag set.  If it was found on the
2240 * unevictable list, it will have the Unevictable flag set.  These flags
2241 * may need to be cleared by the caller before letting the page go.
2242 *
2243 * Context:
2244 *
2245 * (1) Must be called with an elevated refcount on the page. This is a
2246 *     fundamental difference from isolate_lru_pages() (which is called
2247 *     without a stable reference).
2248 * (2) The lru_lock must not be held.
2249 * (3) Interrupts must be enabled.
2250 *
2251 * Return: 0 if the folio was removed from an LRU list.
2252 * -EBUSY if the folio was not on an LRU list.
2253 */
2254int folio_isolate_lru(struct folio *folio)
2255{
2256        int ret = -EBUSY;
2257
2258        VM_BUG_ON_FOLIO(!folio_ref_count(folio), folio);
2259
2260        if (folio_test_clear_lru(folio)) {
2261                struct lruvec *lruvec;
2262
2263                folio_get(folio);
2264                lruvec = folio_lruvec_lock_irq(folio);
2265                lruvec_del_folio(lruvec, folio);
2266                unlock_page_lruvec_irq(lruvec);
2267                ret = 0;
2268        }
2269
2270        return ret;
2271}
2272
2273/*
2274 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
2275 * then get rescheduled. When there are massive number of tasks doing page
2276 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
2277 * the LRU list will go small and be scanned faster than necessary, leading to
2278 * unnecessary swapping, thrashing and OOM.
2279 */
2280static int too_many_isolated(struct pglist_data *pgdat, int file,
2281                struct scan_control *sc)
2282{
2283        unsigned long inactive, isolated;
2284        bool too_many;
2285
2286        if (current_is_kswapd())
2287                return 0;
2288
2289        if (!writeback_throttling_sane(sc))
2290                return 0;
2291
2292        if (file) {
2293                inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
2294                isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
2295        } else {
2296                inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
2297                isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
2298        }
2299
2300        /*
2301         * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
2302         * won't get blocked by normal direct-reclaimers, forming a circular
2303         * deadlock.
2304         */
2305        if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
2306                inactive >>= 3;
2307
2308        too_many = isolated > inactive;
2309
2310        /* Wake up tasks throttled due to too_many_isolated. */
2311        if (!too_many)
2312                wake_throttle_isolated(pgdat);
2313
2314        return too_many;
2315}
2316
2317/*
2318 * move_pages_to_lru() moves folios from private @list to appropriate LRU list.
2319 * On return, @list is reused as a list of folios to be freed by the caller.
2320 *
2321 * Returns the number of pages moved to the given lruvec.
2322 */
2323static unsigned int move_pages_to_lru(struct lruvec *lruvec,
2324                                      struct list_head *list)
2325{
2326        int nr_pages, nr_moved = 0;
2327        LIST_HEAD(folios_to_free);
2328
2329        while (!list_empty(list)) {
2330                struct folio *folio = lru_to_folio(list);
2331
2332                VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
2333                list_del(&folio->lru);
2334                if (unlikely(!folio_evictable(folio))) {
2335                        spin_unlock_irq(&lruvec->lru_lock);
2336                        folio_putback_lru(folio);
2337                        spin_lock_irq(&lruvec->lru_lock);
2338                        continue;
2339                }
2340
2341                /*
2342                 * The folio_set_lru needs to be kept here for list integrity.
2343                 * Otherwise:
2344                 *   #0 move_pages_to_lru             #1 release_pages
2345                 *   if (!folio_put_testzero())
2346                 *                                    if (folio_put_testzero())
2347                 *                                      !lru //skip lru_lock
2348                 *     folio_set_lru()
2349                 *     list_add(&folio->lru,)
2350                 *                                        list_add(&folio->lru,)
2351                 */
2352                folio_set_lru(folio);
2353
2354                if (unlikely(folio_put_testzero(folio))) {
2355                        __folio_clear_lru_flags(folio);
2356
2357                        if (unlikely(folio_test_large(folio))) {
2358                                spin_unlock_irq(&lruvec->lru_lock);
2359                                destroy_large_folio(folio);
2360                                spin_lock_irq(&lruvec->lru_lock);
2361                        } else
2362                                list_add(&folio->lru, &folios_to_free);
2363
2364                        continue;
2365                }
2366
2367                /*
2368                 * All pages were isolated from the same lruvec (and isolation
2369                 * inhibits memcg migration).
2370                 */
2371                VM_BUG_ON_FOLIO(!folio_matches_lruvec(folio, lruvec), folio);
2372                lruvec_add_folio(lruvec, folio);
2373                nr_pages = folio_nr_pages(folio);
2374                nr_moved += nr_pages;
2375                if (folio_test_active(folio))
2376                        workingset_age_nonresident(lruvec, nr_pages);
2377        }
2378
2379        /*
2380         * To save our caller's stack, now use input list for pages to free.
2381         */
2382        list_splice(&folios_to_free, list);
2383
2384        return nr_moved;
2385}
2386
2387/*
2388 * If a kernel thread (such as nfsd for loop-back mounts) services a backing
2389 * device by writing to the page cache it sets PF_LOCAL_THROTTLE. In this case
2390 * we should not throttle.  Otherwise it is safe to do so.
2391 */
2392static int current_may_throttle(void)
2393{
2394        return !(current->flags & PF_LOCAL_THROTTLE);
2395}
2396
2397/*
2398 * shrink_inactive_list() is a helper for shrink_node().  It returns the number
2399 * of reclaimed pages
2400 */
2401static unsigned long
2402shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
2403                     struct scan_control *sc, enum lru_list lru)
2404{
2405        LIST_HEAD(page_list);
2406        unsigned long nr_scanned;
2407        unsigned int nr_reclaimed = 0;
2408        unsigned long nr_taken;
2409        struct reclaim_stat stat;
2410        bool file = is_file_lru(lru);
2411        enum vm_event_item item;
2412        struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2413        bool stalled = false;
2414
2415        while (unlikely(too_many_isolated(pgdat, file, sc))) {
2416                if (stalled)
2417                        return 0;
2418
2419                /* wait a bit for the reclaimer. */
2420                stalled = true;
2421                reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED);
2422
2423                /* We are about to die and free our memory. Return now. */
2424                if (fatal_signal_pending(current))
2425                        return SWAP_CLUSTER_MAX;
2426        }
2427
2428        lru_add_drain();
2429
2430        spin_lock_irq(&lruvec->lru_lock);
2431
2432        nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
2433                                     &nr_scanned, sc, lru);
2434
2435        __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2436        item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
2437        if (!cgroup_reclaim(sc))
2438                __count_vm_events(item, nr_scanned);
2439        __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
2440        __count_vm_events(PGSCAN_ANON + file, nr_scanned);
2441
2442        spin_unlock_irq(&lruvec->lru_lock);
2443
2444        if (nr_taken == 0)
2445                return 0;
2446
2447        nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, &stat, false);
2448
2449        spin_lock_irq(&lruvec->lru_lock);
2450        move_pages_to_lru(lruvec, &page_list);
2451
2452        __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2453        item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
2454        if (!cgroup_reclaim(sc))
2455                __count_vm_events(item, nr_reclaimed);
2456        __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
2457        __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed);
2458        spin_unlock_irq(&lruvec->lru_lock);
2459
2460        lru_note_cost(lruvec, file, stat.nr_pageout);
2461        mem_cgroup_uncharge_list(&page_list);
2462        free_unref_page_list(&page_list);
2463
2464        /*
2465         * If dirty pages are scanned that are not queued for IO, it
2466         * implies that flushers are not doing their job. This can
2467         * happen when memory pressure pushes dirty pages to the end of
2468         * the LRU before the dirty limits are breached and the dirty
2469         * data has expired. It can also happen when the proportion of
2470         * dirty pages grows not through writes but through memory
2471         * pressure reclaiming all the clean cache. And in some cases,
2472         * the flushers simply cannot keep up with the allocation
2473         * rate. Nudge the flusher threads in case they are asleep.
2474         */
2475        if (stat.nr_unqueued_dirty == nr_taken)
2476                wakeup_flusher_threads(WB_REASON_VMSCAN);
2477
2478        sc->nr.dirty += stat.nr_dirty;
2479        sc->nr.congested += stat.nr_congested;
2480        sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2481        sc->nr.writeback += stat.nr_writeback;
2482        sc->nr.immediate += stat.nr_immediate;
2483        sc->nr.taken += nr_taken;
2484        if (file)
2485                sc->nr.file_taken += nr_taken;
2486
2487        trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2488                        nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2489        return nr_reclaimed;
2490}
2491
2492/*
2493 * shrink_active_list() moves folios from the active LRU to the inactive LRU.
2494 *
2495 * We move them the other way if the folio is referenced by one or more
2496 * processes.
2497 *
2498 * If the folios are mostly unmapped, the processing is fast and it is
2499 * appropriate to hold lru_lock across the whole operation.  But if
2500 * the folios are mapped, the processing is slow (folio_referenced()), so
2501 * we should drop lru_lock around each folio.  It's impossible to balance
2502 * this, so instead we remove the folios from the LRU while processing them.
2503 * It is safe to rely on the active flag against the non-LRU folios in here
2504 * because nobody will play with that bit on a non-LRU folio.
2505 *
2506 * The downside is that we have to touch folio->_refcount against each folio.
2507 * But we had to alter folio->flags anyway.
2508 */
2509static void shrink_active_list(unsigned long nr_to_scan,
2510                               struct lruvec *lruvec,
2511                               struct scan_control *sc,
2512                               enum lru_list lru)
2513{
2514        unsigned long nr_taken;
2515        unsigned long nr_scanned;
2516        unsigned long vm_flags;
2517        LIST_HEAD(l_hold);      /* The folios which were snipped off */
2518        LIST_HEAD(l_active);
2519        LIST_HEAD(l_inactive);
2520        unsigned nr_deactivate, nr_activate;
2521        unsigned nr_rotated = 0;
2522        int file = is_file_lru(lru);
2523        struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2524
2525        lru_add_drain();
2526
2527        spin_lock_irq(&lruvec->lru_lock);
2528
2529        nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2530                                     &nr_scanned, sc, lru);
2531
2532        __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2533
2534        if (!cgroup_reclaim(sc))
2535                __count_vm_events(PGREFILL, nr_scanned);
2536        __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2537
2538        spin_unlock_irq(&lruvec->lru_lock);
2539
2540        while (!list_empty(&l_hold)) {
2541                struct folio *folio;
2542
2543                cond_resched();
2544                folio = lru_to_folio(&l_hold);
2545                list_del(&folio->lru);
2546
2547                if (unlikely(!folio_evictable(folio))) {
2548                        folio_putback_lru(folio);
2549                        continue;
2550                }
2551
2552                if (unlikely(buffer_heads_over_limit)) {
2553                        if (folio_test_private(folio) && folio_trylock(folio)) {
2554                                if (folio_test_private(folio))
2555                                        filemap_release_folio(folio, 0);
2556                                folio_unlock(folio);
2557                        }
2558                }
2559
2560                /* Referenced or rmap lock contention: rotate */
2561                if (folio_referenced(folio, 0, sc->target_mem_cgroup,
2562                                     &vm_flags) != 0) {
2563                        /*
2564                         * Identify referenced, file-backed active folios and
2565                         * give them one more trip around the active list. So
2566                         * that executable code get better chances to stay in
2567                         * memory under moderate memory pressure.  Anon folios
2568                         * are not likely to be evicted by use-once streaming
2569                         * IO, plus JVM can create lots of anon VM_EXEC folios,
2570                         * so we ignore them here.
2571                         */
2572                        if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio)) {
2573                                nr_rotated += folio_nr_pages(folio);
2574                                list_add(&folio->lru, &l_active);
2575                                continue;
2576                        }
2577                }
2578
2579                folio_clear_active(folio);      /* we are de-activating */
2580                folio_set_workingset(folio);
2581                list_add(&folio->lru, &l_inactive);
2582        }
2583
2584        /*
2585         * Move folios back to the lru list.
2586         */
2587        spin_lock_irq(&lruvec->lru_lock);
2588
2589        nr_activate = move_pages_to_lru(lruvec, &l_active);
2590        nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
2591        /* Keep all free folios in l_active list */
2592        list_splice(&l_inactive, &l_active);
2593
2594        __count_vm_events(PGDEACTIVATE, nr_deactivate);
2595        __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2596
2597        __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2598        spin_unlock_irq(&lruvec->lru_lock);
2599
2600        mem_cgroup_uncharge_list(&l_active);
2601        free_unref_page_list(&l_active);
2602        trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2603                        nr_deactivate, nr_rotated, sc->priority, file);
2604}
2605
2606static unsigned int reclaim_page_list(struct list_head *page_list,
2607                                      struct pglist_data *pgdat)
2608{
2609        struct reclaim_stat dummy_stat;
2610        unsigned int nr_reclaimed;
2611        struct folio *folio;
2612        struct scan_control sc = {
2613                .gfp_mask = GFP_KERNEL,
2614                .may_writepage = 1,
2615                .may_unmap = 1,
2616                .may_swap = 1,
2617                .no_demotion = 1,
2618        };
2619
2620        nr_reclaimed = shrink_page_list(page_list, pgdat, &sc, &dummy_stat, false);
2621        while (!list_empty(page_list)) {
2622                folio = lru_to_folio(page_list);
2623                list_del(&folio->lru);
2624                folio_putback_lru(folio);
2625        }
2626
2627        return nr_reclaimed;
2628}
2629
2630unsigned long reclaim_pages(struct list_head *folio_list)
2631{
2632        int nid;
2633        unsigned int nr_reclaimed = 0;
2634        LIST_HEAD(node_folio_list);
2635        unsigned int noreclaim_flag;
2636
2637        if (list_empty(folio_list))
2638                return nr_reclaimed;
2639
2640        noreclaim_flag = memalloc_noreclaim_save();
2641
2642        nid = folio_nid(lru_to_folio(folio_list));
2643        do {
2644                struct folio *folio = lru_to_folio(folio_list);
2645
2646                if (nid == folio_nid(folio)) {
2647                        folio_clear_active(folio);
2648                        list_move(&folio->lru, &node_folio_list);
2649                        continue;
2650                }
2651
2652                nr_reclaimed += reclaim_page_list(&node_folio_list, NODE_DATA(nid));
2653                nid = folio_nid(lru_to_folio(folio_list));
2654        } while (!list_empty(folio_list));
2655
2656        nr_reclaimed += reclaim_page_list(&node_folio_list, NODE_DATA(nid));
2657
2658        memalloc_noreclaim_restore(noreclaim_flag);
2659
2660        return nr_reclaimed;
2661}
2662
2663static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2664                                 struct lruvec *lruvec, struct scan_control *sc)
2665{
2666        if (is_active_lru(lru)) {
2667                if (sc->may_deactivate & (1 << is_file_lru(lru)))
2668                        shrink_active_list(nr_to_scan, lruvec, sc, lru);
2669                else
2670                        sc->skipped_deactivate = 1;
2671                return 0;
2672        }
2673
2674        return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2675}
2676
2677/*
2678 * The inactive anon list should be small enough that the VM never has
2679 * to do too much work.
2680 *
2681 * The inactive file list should be small enough to leave most memory
2682 * to the established workingset on the scan-resistant active list,
2683 * but large enough to avoid thrashing the aggregate readahead window.
2684 *
2685 * Both inactive lists should also be large enough that each inactive
2686 * page has a chance to be referenced again before it is reclaimed.
2687 *
2688 * If that fails and refaulting is observed, the inactive list grows.
2689 *
2690 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2691 * on this LRU, maintained by the pageout code. An inactive_ratio
2692 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2693 *
2694 * total     target    max
2695 * memory    ratio     inactive
2696 * -------------------------------------
2697 *   10MB       1         5MB
2698 *  100MB       1        50MB
2699 *    1GB       3       250MB
2700 *   10GB      10       0.9GB
2701 *  100GB      31         3GB
2702 *    1TB     101        10GB
2703 *   10TB     320        32GB
2704 */
2705static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
2706{
2707        enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
2708        unsigned long inactive, active;
2709        unsigned long inactive_ratio;
2710        unsigned long gb;
2711
2712        inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2713        active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
2714
2715        gb = (inactive + active) >> (30 - PAGE_SHIFT);
2716        if (gb)
2717                inactive_ratio = int_sqrt(10 * gb);
2718        else
2719                inactive_ratio = 1;
2720
2721        return inactive * inactive_ratio < active;
2722}
2723
2724enum scan_balance {
2725        SCAN_EQUAL,
2726        SCAN_FRACT,
2727        SCAN_ANON,
2728        SCAN_FILE,
2729};
2730
2731/*
2732 * Determine how aggressively the anon and file LRU lists should be
2733 * scanned.
2734 *
2735 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2736 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2737 */
2738static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2739                           unsigned long *nr)
2740{
2741        struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2742        struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2743        unsigned long anon_cost, file_cost, total_cost;
2744        int swappiness = mem_cgroup_swappiness(memcg);
2745        u64 fraction[ANON_AND_FILE];
2746        u64 denominator = 0;    /* gcc */
2747        enum scan_balance scan_balance;
2748        unsigned long ap, fp;
2749        enum lru_list lru;
2750
2751        /* If we have no swap space, do not bother scanning anon pages. */
2752        if (!sc->may_swap || !can_reclaim_anon_pages(memcg, pgdat->node_id, sc)) {
2753                scan_balance = SCAN_FILE;
2754                goto out;
2755        }
2756
2757        /*
2758         * Global reclaim will swap to prevent OOM even with no
2759         * swappiness, but memcg users want to use this knob to
2760         * disable swapping for individual groups completely when
2761         * using the memory controller's swap limit feature would be
2762         * too expensive.
2763         */
2764        if (cgroup_reclaim(sc) && !swappiness) {
2765                scan_balance = SCAN_FILE;
2766                goto out;
2767        }
2768
2769        /*
2770         * Do not apply any pressure balancing cleverness when the
2771         * system is close to OOM, scan both anon and file equally
2772         * (unless the swappiness setting disagrees with swapping).
2773         */
2774        if (!sc->priority && swappiness) {
2775                scan_balance = SCAN_EQUAL;
2776                goto out;
2777        }
2778
2779        /*
2780         * If the system is almost out of file pages, force-scan anon.
2781         */
2782        if (sc->file_is_tiny) {
2783                scan_balance = SCAN_ANON;
2784                goto out;
2785        }
2786
2787        /*
2788         * If there is enough inactive page cache, we do not reclaim
2789         * anything from the anonymous working right now.
2790         */
2791        if (sc->cache_trim_mode) {
2792                scan_balance = SCAN_FILE;
2793                goto out;
2794        }
2795
2796        scan_balance = SCAN_FRACT;
2797        /*
2798         * Calculate the pressure balance between anon and file pages.
2799         *
2800         * The amount of pressure we put on each LRU is inversely
2801         * proportional to the cost of reclaiming each list, as
2802         * determined by the share of pages that are refaulting, times
2803         * the relative IO cost of bringing back a swapped out
2804         * anonymous page vs reloading a filesystem page (swappiness).
2805         *
2806         * Although we limit that influence to ensure no list gets
2807         * left behind completely: at least a third of the pressure is
2808         * applied, before swappiness.
2809         *
2810         * With swappiness at 100, anon and file have equal IO cost.
2811         */
2812        total_cost = sc->anon_cost + sc->file_cost;
2813        anon_cost = total_cost + sc->anon_cost;
2814        file_cost = total_cost + sc->file_cost;
2815        total_cost = anon_cost + file_cost;
2816
2817        ap = swappiness * (total_cost + 1);
2818        ap /= anon_cost + 1;
2819
2820        fp = (200 - swappiness) * (total_cost + 1);
2821        fp /= file_cost + 1;
2822
2823        fraction[0] = ap;
2824        fraction[1] = fp;
2825        denominator = ap + fp;
2826out:
2827        for_each_evictable_lru(lru) {
2828                int file = is_file_lru(lru);
2829                unsigned long lruvec_size;
2830                unsigned long low, min;
2831                unsigned long scan;
2832
2833                lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2834                mem_cgroup_protection(sc->target_mem_cgroup, memcg,
2835                                      &min, &low);
2836
2837                if (min || low) {
2838                        /*
2839                         * Scale a cgroup's reclaim pressure by proportioning
2840                         * its current usage to its memory.low or memory.min
2841                         * setting.
2842                         *
2843                         * This is important, as otherwise scanning aggression
2844                         * becomes extremely binary -- from nothing as we
2845                         * approach the memory protection threshold, to totally
2846                         * nominal as we exceed it.  This results in requiring
2847                         * setting extremely liberal protection thresholds. It
2848                         * also means we simply get no protection at all if we
2849                         * set it too low, which is not ideal.
2850                         *
2851                         * If there is any protection in place, we reduce scan
2852                         * pressure by how much of the total memory used is
2853                         * within protection thresholds.
2854                         *
2855                         * There is one special case: in the first reclaim pass,
2856                         * we skip over all groups that are within their low
2857                         * protection. If that fails to reclaim enough pages to
2858                         * satisfy the reclaim goal, we come back and override
2859                         * the best-effort low protection. However, we still
2860                         * ideally want to honor how well-behaved groups are in
2861                         * that case instead of simply punishing them all
2862                         * equally. As such, we reclaim them based on how much
2863                         * memory they are using, reducing the scan pressure
2864                         * again by how much of the total memory used is under
2865                         * hard protection.
2866                         */
2867                        unsigned long cgroup_size = mem_cgroup_size(memcg);
2868                        unsigned long protection;
2869
2870                        /* memory.low scaling, make sure we retry before OOM */
2871                        if (!sc->memcg_low_reclaim && low > min) {
2872                                protection = low;
2873                                sc->memcg_low_skipped = 1;
2874                        } else {
2875                                protection = min;
2876                        }
2877
2878                        /* Avoid TOCTOU with earlier protection check */
2879                        cgroup_size = max(cgroup_size, protection);
2880
2881                        scan = lruvec_size - lruvec_size * protection /
2882                                (cgroup_size + 1);
2883
2884                        /*
2885                         * Minimally target SWAP_CLUSTER_MAX pages to keep
2886                         * reclaim moving forwards, avoiding decrementing
2887                         * sc->priority further than desirable.
2888                         */
2889                        scan = max(scan, SWAP_CLUSTER_MAX);
2890                } else {
2891                        scan = lruvec_size;
2892                }
2893
2894                scan >>= sc->priority;
2895
2896                /*
2897                 * If the cgroup's already been deleted, make sure to
2898                 * scrape out the remaining cache.
2899                 */
2900                if (!scan && !mem_cgroup_online(memcg))
2901                        scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2902
2903                switch (scan_balance) {
2904                case SCAN_EQUAL:
2905                        /* Scan lists relative to size */
2906                        break;
2907                case SCAN_FRACT:
2908                        /*
2909                         * Scan types proportional to swappiness and
2910                         * their relative recent reclaim efficiency.
2911                         * Make sure we don't miss the last page on
2912                         * the offlined memory cgroups because of a
2913                         * round-off error.
2914                         */
2915                        scan = mem_cgroup_online(memcg) ?
2916                               div64_u64(scan * fraction[file], denominator) :
2917                               DIV64_U64_ROUND_UP(scan * fraction[file],
2918                                                  denominator);
2919                        break;
2920                case SCAN_FILE:
2921                case SCAN_ANON:
2922                        /* Scan one type exclusively */
2923                        if ((scan_balance == SCAN_FILE) != file)
2924                                scan = 0;
2925                        break;
2926                default:
2927                        /* Look ma, no brain */
2928                        BUG();
2929                }
2930
2931                nr[lru] = scan;
2932        }
2933}
2934
2935/*
2936 * Anonymous LRU management is a waste if there is
2937 * ultimately no way to reclaim the memory.
2938 */
2939static bool can_age_anon_pages(struct pglist_data *pgdat,
2940                               struct scan_control *sc)
2941{
2942        /* Aging the anon LRU is valuable if swap is present: */
2943        if (total_swap_pages > 0)
2944                return true;
2945
2946        /* Also valuable if anon pages can be demoted: */
2947        return can_demote(pgdat->node_id, sc);
2948}
2949
2950static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
2951{
2952        unsigned long nr[NR_LRU_LISTS];
2953        unsigned long targets[NR_LRU_LISTS];
2954        unsigned long nr_to_scan;
2955        enum lru_list lru;
2956        unsigned long nr_reclaimed = 0;
2957        unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2958        struct blk_plug plug;
2959        bool scan_adjusted;
2960
2961        get_scan_count(lruvec, sc, nr);
2962
2963        /* Record the original scan target for proportional adjustments later */
2964        memcpy(targets, nr, sizeof(nr));
2965
2966        /*
2967         * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2968         * event that can occur when there is little memory pressure e.g.
2969         * multiple streaming readers/writers. Hence, we do not abort scanning
2970         * when the requested number of pages are reclaimed when scanning at
2971         * DEF_PRIORITY on the assumption that the fact we are direct
2972         * reclaiming implies that kswapd is not keeping up and it is best to
2973         * do a batch of work at once. For memcg reclaim one check is made to
2974         * abort proportional reclaim if either the file or anon lru has already
2975         * dropped to zero at the first pass.
2976         */
2977        scan_adjusted = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
2978                         sc->priority == DEF_PRIORITY);
2979
2980        blk_start_plug(&plug);
2981        while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2982                                        nr[LRU_INACTIVE_FILE]) {
2983                unsigned long nr_anon, nr_file, percentage;
2984                unsigned long nr_scanned;
2985
2986                for_each_evictable_lru(lru) {
2987                        if (nr[lru]) {
2988                                nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2989                                nr[lru] -= nr_to_scan;
2990
2991                                nr_reclaimed += shrink_list(lru, nr_to_scan,
2992                                                            lruvec, sc);
2993                        }
2994                }
2995
2996                cond_resched();
2997
2998                if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2999                        continue;
3000
3001                /*
3002                 * For kswapd and memcg, reclaim at least the number of pages
3003                 * requested. Ensure that the anon and file LRUs are scanned
3004                 * proportionally what was requested by get_scan_count(). We
3005                 * stop reclaiming one LRU and reduce the amount scanning
3006                 * proportional to the original scan target.
3007                 */
3008                nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
3009                nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
3010
3011                /*
3012                 * It's just vindictive to attack the larger once the smaller
3013                 * has gone to zero.  And given the way we stop scanning the
3014                 * smaller below, this makes sure that we only make one nudge
3015                 * towards proportionality once we've got nr_to_reclaim.
3016                 */
3017                if (!nr_file || !nr_anon)
3018                        break;
3019
3020                if (nr_file > nr_anon) {
3021                        unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
3022                                                targets[LRU_ACTIVE_ANON] + 1;
3023                        lru = LRU_BASE;
3024                        percentage = nr_anon * 100 / scan_target;
3025                } else {
3026                        unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
3027                                                targets[LRU_ACTIVE_FILE] + 1;
3028                        lru = LRU_FILE;
3029                        percentage = nr_file * 100 / scan_target;
3030                }
3031
3032                /* Stop scanning the smaller of the LRU */
3033                nr[lru] = 0;
3034                nr[lru + LRU_ACTIVE] = 0;
3035
3036                /*
3037                 * Recalculate the other LRU scan count based on its original
3038                 * scan target and the percentage scanning already complete
3039                 */
3040                lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
3041                nr_scanned = targets[lru] - nr[lru];
3042                nr[lru] = targets[lru] * (100 - percentage) / 100;
3043                nr[lru] -= min(nr[lru], nr_scanned);
3044
3045                lru += LRU_ACTIVE;
3046                nr_scanned = targets[lru] - nr[lru];
3047                nr[lru] = targets[lru] * (100 - percentage) / 100;
3048                nr[lru] -= min(nr[lru], nr_scanned);
3049
3050                scan_adjusted = true;
3051        }
3052        blk_finish_plug(&plug);
3053        sc->nr_reclaimed += nr_reclaimed;
3054
3055        /*
3056         * Even if we did not try to evict anon pages at all, we want to
3057         * rebalance the anon lru active/inactive ratio.
3058         */
3059        if (can_age_anon_pages(lruvec_pgdat(lruvec), sc) &&
3060            inactive_is_low(lruvec, LRU_INACTIVE_ANON))
3061                shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3062                                   sc, LRU_ACTIVE_ANON);
3063}
3064
3065/* Use reclaim/compaction for costly allocs or under memory pressure */
3066static bool in_reclaim_compaction(struct scan_control *sc)
3067{
3068        if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
3069                        (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
3070                         sc->priority < DEF_PRIORITY - 2))
3071                return true;
3072
3073        return false;
3074}
3075
3076/*
3077 * Reclaim/compaction is used for high-order allocation requests. It reclaims
3078 * order-0 pages before compacting the zone. should_continue_reclaim() returns
3079 * true if more pages should be reclaimed such that when the page allocator
3080 * calls try_to_compact_pages() that it will have enough free pages to succeed.
3081 * It will give up earlier than that if there is difficulty reclaiming pages.
3082 */
3083static inline bool should_continue_reclaim(struct pglist_data *pgdat,
3084                                        unsigned long nr_reclaimed,
3085                                        struct scan_control *sc)
3086{
3087        unsigned long pages_for_compaction;
3088        unsigned long inactive_lru_pages;
3089        int z;
3090
3091        /* If not in reclaim/compaction mode, stop */
3092        if (!in_reclaim_compaction(sc))
3093                return false;
3094
3095        /*
3096         * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
3097         * number of pages that were scanned. This will return to the caller
3098         * with the risk reclaim/compaction and the resulting allocation attempt
3099         * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
3100         * allocations through requiring that the full LRU list has been scanned
3101         * first, by assuming that zero delta of sc->nr_scanned means full LRU
3102         * scan, but that approximation was wrong, and there were corner cases
3103         * where always a non-zero amount of pages were scanned.
3104         */
3105        if (!nr_reclaimed)
3106                return false;
3107
3108        /* If compaction would go ahead or the allocation would succeed, stop */
3109        for (z = 0; z <= sc->reclaim_idx; z++) {
3110                struct zone *zone = &pgdat->node_zones[z];
3111                if (!managed_zone(zone))
3112                        continue;
3113
3114                switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
3115                case COMPACT_SUCCESS:
3116                case COMPACT_CONTINUE:
3117                        return false;
3118                default:
3119                        /* check next zone */
3120                        ;
3121                }
3122        }
3123
3124        /*
3125         * If we have not reclaimed enough pages for compaction and the
3126         * inactive lists are large enough, continue reclaiming
3127         */
3128        pages_for_compaction = compact_gap(sc->order);
3129        inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
3130        if (can_reclaim_anon_pages(NULL, pgdat->node_id, sc))
3131                inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
3132
3133        return inactive_lru_pages > pages_for_compaction;
3134}
3135
3136static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
3137{
3138        struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
3139        struct mem_cgroup *memcg;
3140
3141        memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
3142        do {
3143                struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3144                unsigned long reclaimed;
3145                unsigned long scanned;
3146
3147                /*
3148                 * This loop can become CPU-bound when target memcgs
3149                 * aren't eligible for reclaim - either because they
3150                 * don't have any reclaimable pages, or because their
3151                 * memory is explicitly protected. Avoid soft lockups.
3152                 */
3153                cond_resched();
3154
3155                mem_cgroup_calculate_protection(target_memcg, memcg);
3156
3157                if (mem_cgroup_below_min(memcg)) {
3158                        /*
3159                         * Hard protection.
3160                         * If there is no reclaimable memory, OOM.
3161                         */
3162                        continue;
3163                } else if (mem_cgroup_below_low(memcg)) {
3164                        /*
3165                         * Soft protection.
3166                         * Respect the protection only as long as
3167                         * there is an unprotected supply
3168                         * of reclaimable memory from other cgroups.
3169                         */
3170                        if (!sc->memcg_low_reclaim) {
3171                                sc->memcg_low_skipped = 1;
3172                                continue;
3173                        }
3174                        memcg_memory_event(memcg, MEMCG_LOW);
3175                }
3176
3177                reclaimed = sc->nr_reclaimed;
3178                scanned = sc->nr_scanned;
3179
3180                shrink_lruvec(lruvec, sc);
3181
3182                shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
3183                            sc->priority);
3184
3185                /* Record the group's reclaim efficiency */
3186                if (!sc->proactive)
3187                        vmpressure(sc->gfp_mask, memcg, false,
3188                                   sc->nr_scanned - scanned,
3189                                   sc->nr_reclaimed - reclaimed);
3190
3191        } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
3192}
3193
3194static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
3195{
3196        struct reclaim_state *reclaim_state = current->reclaim_state;
3197        unsigned long nr_reclaimed, nr_scanned;
3198        struct lruvec *target_lruvec;
3199        bool reclaimable = false;
3200        unsigned long file;
3201
3202        target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
3203
3204again:
3205        /*
3206         * Flush the memory cgroup stats, so that we read accurate per-memcg
3207         * lruvec stats for heuristics.
3208         */
3209        mem_cgroup_flush_stats();
3210
3211        memset(&sc->nr, 0, sizeof(sc->nr));
3212
3213        nr_reclaimed = sc->nr_reclaimed;
3214        nr_scanned = sc->nr_scanned;
3215
3216        /*
3217         * Determine the scan balance between anon and file LRUs.
3218         */
3219        spin_lock_irq(&target_lruvec->lru_lock);
3220        sc->anon_cost = target_lruvec->anon_cost;
3221        sc->file_cost = target_lruvec->file_cost;
3222        spin_unlock_irq(&target_lruvec->lru_lock);
3223
3224        /*
3225         * Target desirable inactive:active list ratios for the anon
3226         * and file LRU lists.
3227         */
3228        if (!sc->force_deactivate) {
3229                unsigned long refaults;
3230
3231                refaults = lruvec_page_state(target_lruvec,
3232                                WORKINGSET_ACTIVATE_ANON);
3233                if (refaults != target_lruvec->refaults[0] ||
3234                        inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
3235                        sc->may_deactivate |= DEACTIVATE_ANON;
3236                else
3237                        sc->may_deactivate &= ~DEACTIVATE_ANON;
3238
3239                /*
3240                 * When refaults are being observed, it means a new
3241                 * workingset is being established. Deactivate to get
3242                 * rid of any stale active pages quickly.
3243                 */
3244                refaults = lruvec_page_state(target_lruvec,
3245                                WORKINGSET_ACTIVATE_FILE);
3246                if (refaults != target_lruvec->refaults[1] ||
3247                    inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
3248                        sc->may_deactivate |= DEACTIVATE_FILE;
3249                else
3250                        sc->may_deactivate &= ~DEACTIVATE_FILE;
3251        } else
3252                sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
3253
3254        /*
3255         * If we have plenty of inactive file pages that aren't
3256         * thrashing, try to reclaim those first before touching
3257         * anonymous pages.
3258         */
3259        file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
3260        if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE))
3261                sc->cache_trim_mode = 1;
3262        else
3263                sc->cache_trim_mode = 0;
3264
3265        /*
3266         * Prevent the reclaimer from falling into the cache trap: as
3267         * cache pages start out inactive, every cache fault will tip
3268         * the scan balance towards the file LRU.  And as the file LRU
3269         * shrinks, so does the window for rotation from references.
3270         * This means we have a runaway feedback loop where a tiny
3271         * thrashing file LRU becomes infinitely more attractive than
3272         * anon pages.  Try to detect this based on file LRU size.
3273         */
3274        if (!cgroup_reclaim(sc)) {
3275                unsigned long total_high_wmark = 0;
3276                unsigned long free, anon;
3277                int z;
3278
3279                free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
3280                file = node_page_state(pgdat, NR_ACTIVE_FILE) +
3281                           node_page_state(pgdat, NR_INACTIVE_FILE);
3282
3283                for (z = 0; z < MAX_NR_ZONES; z++) {
3284                        struct zone *zone = &pgdat->node_zones[z];
3285                        if (!managed_zone(zone))
3286                                continue;
3287
3288                        total_high_wmark += high_wmark_pages(zone);
3289                }
3290
3291                /*
3292                 * Consider anon: if that's low too, this isn't a
3293                 * runaway file reclaim problem, but rather just
3294                 * extreme pressure. Reclaim as per usual then.
3295                 */
3296                anon = node_page_state(pgdat, NR_INACTIVE_ANON);
3297
3298                sc->file_is_tiny =
3299                        file + free <= total_high_wmark &&
3300                        !(sc->may_deactivate & DEACTIVATE_ANON) &&
3301                        anon >> sc->priority;
3302        }
3303
3304        shrink_node_memcgs(pgdat, sc);
3305
3306        if (reclaim_state) {
3307                sc->nr_reclaimed += reclaim_state->reclaimed_slab;
3308                reclaim_state->reclaimed_slab = 0;
3309        }
3310
3311        /* Record the subtree's reclaim efficiency */
3312        if (!sc->proactive)
3313                vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
3314                           sc->nr_scanned - nr_scanned,
3315                           sc->nr_reclaimed - nr_reclaimed);
3316
3317        if (sc->nr_reclaimed - nr_reclaimed)
3318                reclaimable = true;
3319
3320        if (current_is_kswapd()) {
3321                /*
3322                 * If reclaim is isolating dirty pages under writeback,
3323                 * it implies that the long-lived page allocation rate
3324                 * is exceeding the page laundering rate. Either the
3325                 * global limits are not being effective at throttling
3326                 * processes due to the page distribution throughout
3327                 * zones or there is heavy usage of a slow backing
3328                 * device. The only option is to throttle from reclaim
3329                 * context which is not ideal as there is no guarantee
3330                 * the dirtying process is throttled in the same way
3331                 * balance_dirty_pages() manages.
3332                 *
3333                 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
3334                 * count the number of pages under pages flagged for
3335                 * immediate reclaim and stall if any are encountered
3336                 * in the nr_immediate check below.
3337                 */
3338                if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
3339                        set_bit(PGDAT_WRITEBACK, &pgdat->flags);
3340
3341                /* Allow kswapd to start writing pages during reclaim.*/
3342                if (sc->nr.unqueued_dirty == sc->nr.file_taken)
3343                        set_bit(PGDAT_DIRTY, &pgdat->flags);
3344
3345                /*
3346                 * If kswapd scans pages marked for immediate
3347                 * reclaim and under writeback (nr_immediate), it
3348                 * implies that pages are cycling through the LRU
3349                 * faster than they are written so forcibly stall
3350                 * until some pages complete writeback.
3351                 */
3352                if (sc->nr.immediate)
3353                        reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK);
3354        }
3355
3356        /*
3357         * Tag a node/memcg as congested if all the dirty pages were marked
3358         * for writeback and immediate reclaim (counted in nr.congested).
3359         *
3360         * Legacy memcg will stall in page writeback so avoid forcibly
3361         * stalling in reclaim_throttle().
3362         */
3363        if ((current_is_kswapd() ||
3364             (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) &&
3365            sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
3366                set_bit(LRUVEC_CONGESTED, &target_lruvec->flags);
3367
3368        /*
3369         * Stall direct reclaim for IO completions if the lruvec is
3370         * node is congested. Allow kswapd to continue until it
3371         * starts encountering unqueued dirty pages or cycling through
3372         * the LRU too quickly.
3373         */
3374        if (!current_is_kswapd() && current_may_throttle() &&
3375            !sc->hibernation_mode &&
3376            test_bit(LRUVEC_CONGESTED, &target_lruvec->flags))
3377                reclaim_throttle(pgdat, VMSCAN_THROTTLE_CONGESTED);
3378
3379        if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
3380                                    sc))
3381                goto again;
3382
3383        /*
3384         * Kswapd gives up on balancing particular nodes after too
3385         * many failures to reclaim anything from them and goes to
3386         * sleep. On reclaim progress, reset the failure counter. A
3387         * successful direct reclaim run will revive a dormant kswapd.
3388         */
3389        if (reclaimable)
3390                pgdat->kswapd_failures = 0;
3391}
3392
3393/*
3394 * Returns true if compaction should go ahead for a costly-order request, or
3395 * the allocation would already succeed without compaction. Return false if we
3396 * should reclaim first.
3397 */
3398static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
3399{
3400        unsigned long watermark;
3401        enum compact_result suitable;
3402
3403        suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
3404        if (suitable == COMPACT_SUCCESS)
3405                /* Allocation should succeed already. Don't reclaim. */
3406                return true;
3407        if (suitable == COMPACT_SKIPPED)
3408                /* Compaction cannot yet proceed. Do reclaim. */
3409                return false;
3410
3411        /*
3412         * Compaction is already possible, but it takes time to run and there
3413         * are potentially other callers using the pages just freed. So proceed
3414         * with reclaim to make a buffer of free pages available to give
3415         * compaction a reasonable chance of completing and allocating the page.
3416         * Note that we won't actually reclaim the whole buffer in one attempt
3417         * as the target watermark in should_continue_reclaim() is lower. But if
3418         * we are already above the high+gap watermark, don't reclaim at all.
3419         */
3420        watermark = high_wmark_pages(zone) + compact_gap(sc->order);
3421
3422        return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
3423}
3424
3425static void consider_reclaim_throttle(pg_data_t *pgdat, struct scan_control *sc)
3426{
3427        /*
3428         * If reclaim is making progress greater than 12% efficiency then
3429         * wake all the NOPROGRESS throttled tasks.
3430         */
3431        if (sc->nr_reclaimed > (sc->nr_scanned >> 3)) {
3432                wait_queue_head_t *wqh;
3433
3434                wqh = &pgdat->reclaim_wait[VMSCAN_THROTTLE_NOPROGRESS];
3435                if (waitqueue_active(wqh))
3436                        wake_up(wqh);
3437
3438                return;
3439        }
3440
3441        /*
3442         * Do not throttle kswapd or cgroup reclaim on NOPROGRESS as it will
3443         * throttle on VMSCAN_THROTTLE_WRITEBACK if there are too many pages
3444         * under writeback and marked for immediate reclaim at the tail of the
3445         * LRU.
3446         */
3447        if (current_is_kswapd() || cgroup_reclaim(sc))
3448                return;
3449
3450        /* Throttle if making no progress at high prioities. */
3451        if (sc->priority == 1 && !sc->nr_reclaimed)
3452                reclaim_throttle(pgdat, VMSCAN_THROTTLE_NOPROGRESS);
3453}
3454
3455/*
3456 * This is the direct reclaim path, for page-allocating processes.  We only
3457 * try to reclaim pages from zones which will satisfy the caller's allocation
3458 * request.
3459 *
3460 * If a zone is deemed to be full of pinned pages then just give it a light
3461 * scan then give up on it.
3462 */
3463static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
3464{
3465        struct zoneref *z;
3466        struct zone *zone;
3467        unsigned long nr_soft_reclaimed;
3468        unsigned long nr_soft_scanned;
3469        gfp_t orig_mask;
3470        pg_data_t *last_pgdat = NULL;
3471        pg_data_t *first_pgdat = NULL;
3472
3473        /*
3474         * If the number of buffer_heads in the machine exceeds the maximum
3475         * allowed level, force direct reclaim to scan the highmem zone as
3476         * highmem pages could be pinning lowmem pages storing buffer_heads
3477         */
3478        orig_mask = sc->gfp_mask;
3479        if (buffer_heads_over_limit) {
3480                sc->gfp_mask |= __GFP_HIGHMEM;
3481                sc->reclaim_idx = gfp_zone(sc->gfp_mask);
3482        }
3483
3484        for_each_zone_zonelist_nodemask(zone, z, zonelist,
3485                                        sc->reclaim_idx, sc->nodemask) {
3486                /*
3487                 * Take care memory controller reclaiming has small influence
3488                 * to global LRU.
3489                 */
3490                if (!cgroup_reclaim(sc)) {
3491                        if (!cpuset_zone_allowed(zone,
3492                                                 GFP_KERNEL | __GFP_HARDWALL))
3493                                continue;
3494
3495                        /*
3496                         * If we already have plenty of memory free for
3497                         * compaction in this zone, don't free any more.
3498                         * Even though compaction is invoked for any
3499                         * non-zero order, only frequent costly order
3500                         * reclamation is disruptive enough to become a
3501                         * noticeable problem, like transparent huge
3502                         * page allocations.
3503                         */
3504                        if (IS_ENABLED(CONFIG_COMPACTION) &&
3505                            sc->order > PAGE_ALLOC_COSTLY_ORDER &&
3506                            compaction_ready(zone, sc)) {
3507                                sc->compaction_ready = true;
3508                                continue;
3509                        }
3510
3511                        /*
3512                         * Shrink each node in the zonelist once. If the
3513                         * zonelist is ordered by zone (not the default) then a
3514                         * node may be shrunk multiple times but in that case
3515                         * the user prefers lower zones being preserved.
3516                         */
3517                        if (zone->zone_pgdat == last_pgdat)
3518                                continue;
3519
3520                        /*
3521                         * This steals pages from memory cgroups over softlimit
3522                         * and returns the number of reclaimed pages and
3523                         * scanned pages. This works for global memory pressure
3524                         * and balancing, not for a memcg's limit.
3525                         */
3526                        nr_soft_scanned = 0;
3527                        nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
3528                                                sc->order, sc->gfp_mask,
3529                                                &nr_soft_scanned);
3530                        sc->nr_reclaimed += nr_soft_reclaimed;
3531                        sc->nr_scanned += nr_soft_scanned;
3532                        /* need some check for avoid more shrink_zone() */
3533                }
3534
3535                if (!first_pgdat)
3536                        first_pgdat = zone->zone_pgdat;
3537
3538                /* See comment about same check for global reclaim above */
3539                if (zone->zone_pgdat == last_pgdat)
3540                        continue;
3541                last_pgdat = zone->zone_pgdat;
3542                shrink_node(zone->zone_pgdat, sc);
3543        }
3544
3545        if (first_pgdat)
3546                consider_reclaim_throttle(first_pgdat, sc);
3547
3548        /*
3549         * Restore to original mask to avoid the impact on the caller if we
3550         * promoted it to __GFP_HIGHMEM.
3551         */
3552        sc->gfp_mask = orig_mask;
3553}
3554
3555static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
3556{
3557        struct lruvec *target_lruvec;
3558        unsigned long refaults;
3559
3560        target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
3561        refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON);
3562        target_lruvec->refaults[0] = refaults;
3563        refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE);
3564        target_lruvec->refaults[1] = refaults;
3565}
3566
3567/*
3568 * This is the main entry point to direct page reclaim.
3569 *
3570 * If a full scan of the inactive list fails to free enough memory then we
3571 * are "out of memory" and something needs to be killed.
3572 *
3573 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3574 * high - the zone may be full of dirty or under-writeback pages, which this
3575 * caller can't do much about.  We kick the writeback threads and take explicit
3576 * naps in the hope that some of these pages can be written.  But if the
3577 * allocating task holds filesystem locks which prevent writeout this might not
3578 * work, and the allocation attempt will fail.
3579 *
3580 * returns:     0, if no pages reclaimed
3581 *              else, the number of pages reclaimed
3582 */
3583static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3584                                          struct scan_control *sc)
3585{
3586        int initial_priority = sc->priority;
3587        pg_data_t *last_pgdat;
3588        struct zoneref *z;
3589        struct zone *zone;
3590retry:
3591        delayacct_freepages_start();
3592
3593        if (!cgroup_reclaim(sc))
3594                __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
3595
3596        do {
3597                if (!sc->proactive)
3598                        vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3599                                        sc->priority);
3600                sc->nr_scanned = 0;
3601                shrink_zones(zonelist, sc);
3602
3603                if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3604                        break;
3605
3606                if (sc->compaction_ready)
3607                        break;
3608
3609                /*
3610                 * If we're getting trouble reclaiming, start doing
3611                 * writepage even in laptop mode.
3612                 */
3613                if (sc->priority < DEF_PRIORITY - 2)
3614                        sc->may_writepage = 1;
3615        } while (--sc->priority >= 0);
3616
3617        last_pgdat = NULL;
3618        for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3619                                        sc->nodemask) {
3620                if (zone->zone_pgdat == last_pgdat)
3621                        continue;
3622                last_pgdat = zone->zone_pgdat;
3623
3624                snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
3625
3626                if (cgroup_reclaim(sc)) {
3627                        struct lruvec *lruvec;
3628
3629                        lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
3630                                                   zone->zone_pgdat);
3631                        clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3632                }
3633        }
3634
3635        delayacct_freepages_end();
3636
3637        if (sc->nr_reclaimed)
3638                return sc->nr_reclaimed;
3639
3640        /* Aborted reclaim to try compaction? don't OOM, then */
3641        if (sc->compaction_ready)
3642                return 1;
3643
3644        /*
3645         * We make inactive:active ratio decisions based on the node's
3646         * composition of memory, but a restrictive reclaim_idx or a
3647         * memory.low cgroup setting can exempt large amounts of
3648         * memory from reclaim. Neither of which are very common, so
3649         * instead of doing costly eligibility calculations of the
3650         * entire cgroup subtree up front, we assume the estimates are
3651         * good, and retry with forcible deactivation if that fails.
3652         */
3653        if (sc->skipped_deactivate) {
3654                sc->priority = initial_priority;
3655                sc->force_deactivate = 1;
3656                sc->skipped_deactivate = 0;
3657                goto retry;
3658        }
3659
3660        /* Untapped cgroup reserves?  Don't OOM, retry. */
3661        if (sc->memcg_low_skipped) {
3662                sc->priority = initial_priority;
3663                sc->force_deactivate = 0;
3664                sc->memcg_low_reclaim = 1;
3665                sc->memcg_low_skipped = 0;
3666                goto retry;
3667        }
3668
3669        return 0;
3670}
3671
3672static bool allow_direct_reclaim(pg_data_t *pgdat)
3673{
3674        struct zone *zone;
3675        unsigned long pfmemalloc_reserve = 0;
3676        unsigned long free_pages = 0;
3677        int i;
3678        bool wmark_ok;
3679
3680        if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3681                return true;
3682
3683        for (i = 0; i <= ZONE_NORMAL; i++) {
3684                zone = &pgdat->node_zones[i];
3685                if (!managed_zone(zone))
3686                        continue;
3687
3688                if (!zone_reclaimable_pages(zone))
3689                        continue;
3690
3691                pfmemalloc_reserve += min_wmark_pages(zone);
3692                free_pages += zone_page_state(zone, NR_FREE_PAGES);
3693        }
3694
3695        /* If there are no reserves (unexpected config) then do not throttle */
3696        if (!pfmemalloc_reserve)
3697                return true;
3698
3699        wmark_ok = free_pages > pfmemalloc_reserve / 2;
3700
3701        /* kswapd must be awake if processes are being throttled */
3702        if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3703                if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
3704                        WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
3705
3706                wake_up_interruptible(&pgdat->kswapd_wait);
3707        }
3708
3709        return wmark_ok;
3710}
3711
3712/*
3713 * Throttle direct reclaimers if backing storage is backed by the network
3714 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3715 * depleted. kswapd will continue to make progress and wake the processes
3716 * when the low watermark is reached.
3717 *
3718 * Returns true if a fatal signal was delivered during throttling. If this
3719 * happens, the page allocator should not consider triggering the OOM killer.
3720 */
3721static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3722                                        nodemask_t *nodemask)
3723{
3724        struct zoneref *z;
3725        struct zone *zone;
3726        pg_data_t *pgdat = NULL;
3727
3728        /*
3729         * Kernel threads should not be throttled as they may be indirectly
3730         * responsible for cleaning pages necessary for reclaim to make forward
3731         * progress. kjournald for example may enter direct reclaim while
3732         * committing a transaction where throttling it could forcing other
3733         * processes to block on log_wait_commit().
3734         */
3735        if (current->flags & PF_KTHREAD)
3736                goto out;
3737
3738        /*
3739         * If a fatal signal is pending, this process should not throttle.
3740         * It should return quickly so it can exit and free its memory
3741         */
3742        if (fatal_signal_pending(current))
3743                goto out;
3744
3745        /*
3746         * Check if the pfmemalloc reserves are ok by finding the first node
3747         * with a usable ZONE_NORMAL or lower zone. The expectation is that
3748         * GFP_KERNEL will be required for allocating network buffers when
3749         * swapping over the network so ZONE_HIGHMEM is unusable.
3750         *
3751         * Throttling is based on the first usable node and throttled processes
3752         * wait on a queue until kswapd makes progress and wakes them. There
3753         * is an affinity then between processes waking up and where reclaim
3754         * progress has been made assuming the process wakes on the same node.
3755         * More importantly, processes running on remote nodes will not compete
3756         * for remote pfmemalloc reserves and processes on different nodes
3757         * should make reasonable progress.
3758         */
3759        for_each_zone_zonelist_nodemask(zone, z, zonelist,
3760                                        gfp_zone(gfp_mask), nodemask) {
3761                if (zone_idx(zone) > ZONE_NORMAL)
3762                        continue;
3763
3764                /* Throttle based on the first usable node */
3765                pgdat = zone->zone_pgdat;
3766                if (allow_direct_reclaim(pgdat))
3767                        goto out;
3768                break;
3769        }
3770
3771        /* If no zone was usable by the allocation flags then do not throttle */
3772        if (!pgdat)
3773                goto out;
3774
3775        /* Account for the throttling */
3776        count_vm_event(PGSCAN_DIRECT_THROTTLE);
3777
3778        /*
3779         * If the caller cannot enter the filesystem, it's possible that it
3780         * is due to the caller holding an FS lock or performing a journal
3781         * transaction in the case of a filesystem like ext[3|4]. In this case,
3782         * it is not safe to block on pfmemalloc_wait as kswapd could be
3783         * blocked waiting on the same lock. Instead, throttle for up to a
3784         * second before continuing.
3785         */
3786        if (!(gfp_mask & __GFP_FS))
3787                wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3788                        allow_direct_reclaim(pgdat), HZ);
3789        else
3790                /* Throttle until kswapd wakes the process */
3791                wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3792                        allow_direct_reclaim(pgdat));
3793
3794        if (fatal_signal_pending(current))
3795                return true;
3796
3797out:
3798        return false;
3799}
3800
3801unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3802                                gfp_t gfp_mask, nodemask_t *nodemask)
3803{
3804        unsigned long nr_reclaimed;
3805        struct scan_control sc = {
3806                .nr_to_reclaim = SWAP_CLUSTER_MAX,
3807                .gfp_mask = current_gfp_context(gfp_mask),
3808                .reclaim_idx = gfp_zone(gfp_mask),
3809                .order = order,
3810                .nodemask = nodemask,
3811                .priority = DEF_PRIORITY,
3812                .may_writepage = !laptop_mode,
3813                .may_unmap = 1,
3814                .may_swap = 1,
3815        };
3816
3817        /*
3818         * scan_control uses s8 fields for order, priority, and reclaim_idx.
3819         * Confirm they are large enough for max values.
3820         */
3821        BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3822        BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3823        BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3824
3825        /*
3826         * Do not enter reclaim if fatal signal was delivered while throttled.
3827         * 1 is returned so that the page allocator does not OOM kill at this
3828         * point.
3829         */
3830        if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3831                return 1;
3832
3833        set_task_reclaim_state(current, &sc.reclaim_state);
3834        trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
3835
3836        nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3837
3838        trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3839        set_task_reclaim_state(current, NULL);
3840
3841        return nr_reclaimed;
3842}
3843
3844#ifdef CONFIG_MEMCG
3845
3846/* Only used by soft limit reclaim. Do not reuse for anything else. */
3847unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3848                                                gfp_t gfp_mask, bool noswap,
3849                                                pg_data_t *pgdat,
3850                                                unsigned long *nr_scanned)
3851{
3852        struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3853        struct scan_control sc = {
3854                .nr_to_reclaim = SWAP_CLUSTER_MAX,
3855                .target_mem_cgroup = memcg,
3856                .may_writepage = !laptop_mode,
3857                .may_unmap = 1,
3858                .reclaim_idx = MAX_NR_ZONES - 1,
3859                .may_swap = !noswap,
3860        };
3861
3862        WARN_ON_ONCE(!current->reclaim_state);
3863
3864        sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3865                        (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3866
3867        trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3868                                                      sc.gfp_mask);
3869
3870        /*
3871         * NOTE: Although we can get the priority field, using it
3872         * here is not a good idea, since it limits the pages we can scan.
3873         * if we don't reclaim here, the shrink_node from balance_pgdat
3874         * will pick up pages from other mem cgroup's as well. We hack
3875         * the priority and make it zero.
3876         */
3877        shrink_lruvec(lruvec, &sc);
3878
3879        trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3880
3881        *nr_scanned = sc.nr_scanned;
3882
3883        return sc.nr_reclaimed;
3884}
3885
3886unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3887                                           unsigned long nr_pages,
3888                                           gfp_t gfp_mask,
3889                                           unsigned int reclaim_options)
3890{
3891        unsigned long nr_reclaimed;
3892        unsigned int noreclaim_flag;
3893        struct scan_control sc = {
3894                .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3895                .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3896                                (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3897                .reclaim_idx = MAX_NR_ZONES - 1,
3898                .target_mem_cgroup = memcg,
3899                .priority = DEF_PRIORITY,
3900                .may_writepage = !laptop_mode,
3901                .may_unmap = 1,
3902                .may_swap = !!(reclaim_options & MEMCG_RECLAIM_MAY_SWAP),
3903                .proactive = !!(reclaim_options & MEMCG_RECLAIM_PROACTIVE),
3904        };
3905        /*
3906         * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
3907         * equal pressure on all the nodes. This is based on the assumption that
3908         * the reclaim does not bail out early.
3909         */
3910        struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3911
3912        set_task_reclaim_state(current, &sc.reclaim_state);
3913        trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
3914        noreclaim_flag = memalloc_noreclaim_save();
3915
3916        nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3917
3918        memalloc_noreclaim_restore(noreclaim_flag);
3919        trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3920        set_task_reclaim_state(current, NULL);
3921
3922        return nr_reclaimed;
3923}
3924#endif
3925
3926static void age_active_anon(struct pglist_data *pgdat,
3927                                struct scan_control *sc)
3928{
3929        struct mem_cgroup *memcg;
3930        struct lruvec *lruvec;
3931
3932        if (!can_age_anon_pages(pgdat, sc))
3933                return;
3934
3935        lruvec = mem_cgroup_lruvec(NULL, pgdat);
3936        if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
3937                return;
3938
3939        memcg = mem_cgroup_iter(NULL, NULL, NULL);
3940        do {
3941                lruvec = mem_cgroup_lruvec(memcg, pgdat);
3942                shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3943                                   sc, LRU_ACTIVE_ANON);
3944                memcg = mem_cgroup_iter(NULL, memcg, NULL);
3945        } while (memcg);
3946}
3947
3948static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
3949{
3950        int i;
3951        struct zone *zone;
3952
3953        /*
3954         * Check for watermark boosts top-down as the higher zones
3955         * are more likely to be boosted. Both watermarks and boosts
3956         * should not be checked at the same time as reclaim would
3957         * start prematurely when there is no boosting and a lower
3958         * zone is balanced.
3959         */
3960        for (i = highest_zoneidx; i >= 0; i--) {
3961                zone = pgdat->node_zones + i;
3962                if (!managed_zone(zone))
3963                        continue;
3964
3965                if (zone->watermark_boost)
3966                        return true;
3967        }
3968
3969        return false;
3970}
3971
3972/*
3973 * Returns true if there is an eligible zone balanced for the request order
3974 * and highest_zoneidx
3975 */
3976static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
3977{
3978        int i;
3979        unsigned long mark = -1;
3980        struct zone *zone;
3981
3982        /*
3983         * Check watermarks bottom-up as lower zones are more likely to
3984         * meet watermarks.
3985         */
3986        for (i = 0; i <= highest_zoneidx; i++) {
3987                zone = pgdat->node_zones + i;
3988
3989                if (!managed_zone(zone))
3990                        continue;
3991
3992                if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING)
3993                        mark = wmark_pages(zone, WMARK_PROMO);
3994                else
3995                        mark = high_wmark_pages(zone);
3996                if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx))
3997                        return true;
3998        }
3999
4000        /*
4001         * If a node has no managed zone within highest_zoneidx, it does not
4002         * need balancing by definition. This can happen if a zone-restricted
4003         * allocation tries to wake a remote kswapd.
4004         */
4005        if (mark == -1)
4006                return true;
4007
4008        return false;
4009}
4010
4011/* Clear pgdat state for congested, dirty or under writeback. */
4012static void clear_pgdat_congested(pg_data_t *pgdat)
4013{
4014        struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
4015
4016        clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
4017        clear_bit(PGDAT_DIRTY, &pgdat->flags);
4018        clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
4019}
4020
4021/*
4022 * Prepare kswapd for sleeping. This verifies that there are no processes
4023 * waiting in throttle_direct_reclaim() and that watermarks have been met.
4024 *
4025 * Returns true if kswapd is ready to sleep
4026 */
4027static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
4028                                int highest_zoneidx)
4029{
4030        /*
4031         * The throttled processes are normally woken up in balance_pgdat() as
4032         * soon as allow_direct_reclaim() is true. But there is a potential
4033         * race between when kswapd checks the watermarks and a process gets
4034         * throttled. There is also a potential race if processes get
4035         * throttled, kswapd wakes, a large process exits thereby balancing the
4036         * zones, which causes kswapd to exit balance_pgdat() before reaching
4037         * the wake up checks. If kswapd is going to sleep, no process should
4038         * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
4039         * the wake up is premature, processes will wake kswapd and get
4040         * throttled again. The difference from wake ups in balance_pgdat() is
4041         * that here we are under prepare_to_wait().
4042         */
4043        if (waitqueue_active(&pgdat->pfmemalloc_wait))
4044                wake_up_all(&pgdat->pfmemalloc_wait);
4045
4046        /* Hopeless node, leave it to direct reclaim */
4047        if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
4048                return true;
4049
4050        if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
4051                clear_pgdat_congested(pgdat);
4052                return true;
4053        }
4054
4055        return false;
4056}
4057
4058/*
4059 * kswapd shrinks a node of pages that are at or below the highest usable
4060 * zone that is currently unbalanced.
4061 *
4062 * Returns true if kswapd scanned at least the requested number of pages to
4063 * reclaim or if the lack of progress was due to pages under writeback.
4064 * This is used to determine if the scanning priority needs to be raised.
4065 */
4066static bool kswapd_shrink_node(pg_data_t *pgdat,
4067                               struct scan_control *sc)
4068{
4069        struct zone *zone;
4070        int z;
4071
4072        /* Reclaim a number of pages proportional to the number of zones */
4073        sc->nr_to_reclaim = 0;
4074        for (z = 0; z <= sc->reclaim_idx; z++) {
4075                zone = pgdat->node_zones + z;
4076                if (!managed_zone(zone))
4077                        continue;
4078
4079                sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
4080        }
4081
4082        /*
4083         * Historically care was taken to put equal pressure on all zones but
4084         * now pressure is applied based on node LRU order.
4085         */
4086        shrink_node(pgdat, sc);
4087
4088        /*
4089         * Fragmentation may mean that the system cannot be rebalanced for
4090         * high-order allocations. If twice the allocation size has been
4091         * reclaimed then recheck watermarks only at order-0 to prevent
4092         * excessive reclaim. Assume that a process requested a high-order
4093         * can direct reclaim/compact.
4094         */
4095        if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
4096                sc->order = 0;
4097
4098        return sc->nr_scanned >= sc->nr_to_reclaim;
4099}
4100
4101/* Page allocator PCP high watermark is lowered if reclaim is active. */
4102static inline void
4103update_reclaim_active(pg_data_t *pgdat, int highest_zoneidx, bool active)
4104{
4105        int i;
4106        struct zone *zone;
4107
4108        for (i = 0; i <= highest_zoneidx; i++) {
4109                zone = pgdat->node_zones + i;
4110
4111                if (!managed_zone(zone))
4112                        continue;
4113
4114                if (active)
4115                        set_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
4116                else
4117                        clear_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
4118        }
4119}
4120
4121static inline void
4122set_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
4123{
4124        update_reclaim_active(pgdat, highest_zoneidx, true);
4125}
4126
4127static inline void
4128clear_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
4129{
4130        update_reclaim_active(pgdat, highest_zoneidx, false);
4131}
4132
4133/*
4134 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
4135 * that are eligible for use by the caller until at least one zone is
4136 * balanced.
4137 *
4138 * Returns the order kswapd finished reclaiming at.
4139 *
4140 * kswapd scans the zones in the highmem->normal->dma direction.  It skips
4141 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
4142 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
4143 * or lower is eligible for reclaim until at least one usable zone is
4144 * balanced.
4145 */
4146static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
4147{
4148        int i;
4149        unsigned long nr_soft_reclaimed;
4150        unsigned long nr_soft_scanned;
4151        unsigned long pflags;
4152        unsigned long nr_boost_reclaim;
4153        unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
4154        bool boosted;
4155        struct zone *zone;
4156        struct scan_control sc = {
4157                .gfp_mask = GFP_KERNEL,
4158                .order = order,
4159                .may_unmap = 1,
4160        };
4161
4162        set_task_reclaim_state(current, &sc.reclaim_state);
4163        psi_memstall_enter(&pflags);
4164        __fs_reclaim_acquire(_THIS_IP_);
4165
4166        count_vm_event(PAGEOUTRUN);
4167
4168        /*
4169         * Account for the reclaim boost. Note that the zone boost is left in
4170         * place so that parallel allocations that are near the watermark will
4171         * stall or direct reclaim until kswapd is finished.
4172         */
4173        nr_boost_reclaim = 0;
4174        for (i = 0; i <= highest_zoneidx; i++) {
4175                zone = pgdat->node_zones + i;
4176                if (!managed_zone(zone))
4177                        continue;
4178
4179                nr_boost_reclaim += zone->watermark_boost;
4180                zone_boosts[i] = zone->watermark_boost;
4181        }
4182        boosted = nr_boost_reclaim;
4183
4184restart:
4185        set_reclaim_active(pgdat, highest_zoneidx);
4186        sc.priority = DEF_PRIORITY;
4187        do {
4188                unsigned long nr_reclaimed = sc.nr_reclaimed;
4189                bool raise_priority = true;
4190                bool balanced;
4191                bool ret;
4192
4193                sc.reclaim_idx = highest_zoneidx;
4194
4195                /*
4196                 * If the number of buffer_heads exceeds the maximum allowed
4197                 * then consider reclaiming from all zones. This has a dual
4198                 * purpose -- on 64-bit systems it is expected that
4199                 * buffer_heads are stripped during active rotation. On 32-bit
4200                 * systems, highmem pages can pin lowmem memory and shrinking
4201                 * buffers can relieve lowmem pressure. Reclaim may still not
4202                 * go ahead if all eligible zones for the original allocation
4203                 * request are balanced to avoid excessive reclaim from kswapd.
4204                 */
4205                if (buffer_heads_over_limit) {
4206                        for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
4207                                zone = pgdat->node_zones + i;
4208                                if (!managed_zone(zone))
4209                                        continue;
4210
4211                                sc.reclaim_idx = i;
4212                                break;
4213                        }
4214                }
4215
4216                /*
4217                 * If the pgdat is imbalanced then ignore boosting and preserve
4218                 * the watermarks for a later time and restart. Note that the
4219                 * zone watermarks will be still reset at the end of balancing
4220                 * on the grounds that the normal reclaim should be enough to
4221                 * re-evaluate if boosting is required when kswapd next wakes.
4222                 */
4223                balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx);
4224                if (!balanced && nr_boost_reclaim) {
4225                        nr_boost_reclaim = 0;
4226                        goto restart;
4227                }
4228
4229                /*
4230                 * If boosting is not active then only reclaim if there are no
4231                 * eligible zones. Note that sc.reclaim_idx is not used as
4232                 * buffer_heads_over_limit may have adjusted it.
4233                 */
4234                if (!nr_boost_reclaim && balanced)
4235                        goto out;
4236
4237                /* Limit the priority of boosting to avoid reclaim writeback */
4238                if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
4239                        raise_priority = false;
4240
4241                /*
4242                 * Do not writeback or swap pages for boosted reclaim. The
4243                 * intent is to relieve pressure not issue sub-optimal IO
4244                 * from reclaim context. If no pages are reclaimed, the
4245                 * reclaim will be aborted.
4246                 */
4247                sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
4248                sc.may_swap = !nr_boost_reclaim;
4249
4250                /*
4251                 * Do some background aging of the anon list, to give
4252                 * pages a chance to be referenced before reclaiming. All
4253                 * pages are rotated regardless of classzone as this is
4254                 * about consistent aging.
4255                 */
4256                age_active_anon(pgdat, &sc);
4257
4258                /*
4259                 * If we're getting trouble reclaiming, start doing writepage
4260                 * even in laptop mode.
4261                 */
4262                if (sc.priority < DEF_PRIORITY - 2)
4263                        sc.may_writepage = 1;
4264
4265                /* Call soft limit reclaim before calling shrink_node. */
4266                sc.nr_scanned = 0;
4267                nr_soft_scanned = 0;
4268                nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
4269                                                sc.gfp_mask, &nr_soft_scanned);
4270                sc.nr_reclaimed += nr_soft_reclaimed;
4271
4272                /*
4273                 * There should be no need to raise the scanning priority if
4274                 * enough pages are already being scanned that that high
4275                 * watermark would be met at 100% efficiency.
4276                 */
4277                if (kswapd_shrink_node(pgdat, &sc))
4278                        raise_priority = false;
4279
4280                /*
4281                 * If the low watermark is met there is no need for processes
4282                 * to be throttled on pfmemalloc_wait as they should not be
4283                 * able to safely make forward progress. Wake them
4284                 */
4285                if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
4286                                allow_direct_reclaim(pgdat))
4287                        wake_up_all(&pgdat->pfmemalloc_wait);
4288
4289                /* Check if kswapd should be suspending */
4290                __fs_reclaim_release(_THIS_IP_);
4291                ret = try_to_freeze();
4292                __fs_reclaim_acquire(_THIS_IP_);
4293                if (ret || kthread_should_stop())
4294                        break;
4295
4296                /*
4297                 * Raise priority if scanning rate is too low or there was no
4298                 * progress in reclaiming pages
4299                 */
4300                nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
4301                nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
4302
4303                /*
4304                 * If reclaim made no progress for a boost, stop reclaim as
4305                 * IO cannot be queued and it could be an infinite loop in
4306                 * extreme circumstances.
4307                 */
4308                if (nr_boost_reclaim && !nr_reclaimed)
4309                        break;
4310
4311                if (raise_priority || !nr_reclaimed)
4312                        sc.priority--;
4313        } while (sc.priority >= 1);
4314
4315        if (!sc.nr_reclaimed)
4316                pgdat->kswapd_failures++;
4317
4318out:
4319        clear_reclaim_active(pgdat, highest_zoneidx);
4320
4321        /* If reclaim was boosted, account for the reclaim done in this pass */
4322        if (boosted) {
4323                unsigned long flags;
4324
4325                for (i = 0; i <= highest_zoneidx; i++) {
4326                        if (!zone_boosts[i])
4327                                continue;
4328
4329                        /* Increments are under the zone lock */
4330                        zone = pgdat->node_zones + i;
4331                        spin_lock_irqsave(&zone->lock, flags);
4332                        zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
4333                        spin_unlock_irqrestore(&zone->lock, flags);
4334                }
4335
4336                /*
4337                 * As there is now likely space, wakeup kcompact to defragment
4338                 * pageblocks.
4339                 */
4340                wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
4341        }
4342
4343        snapshot_refaults(NULL, pgdat);
4344        __fs_reclaim_release(_THIS_IP_);
4345        psi_memstall_leave(&pflags);
4346        set_task_reclaim_state(current, NULL);
4347
4348        /*
4349         * Return the order kswapd stopped reclaiming at as
4350         * prepare_kswapd_sleep() takes it into account. If another caller
4351         * entered the allocator slow path while kswapd was awake, order will
4352         * remain at the higher level.
4353         */
4354        return sc.order;
4355}
4356
4357/*
4358 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
4359 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
4360 * not a valid index then either kswapd runs for first time or kswapd couldn't
4361 * sleep after previous reclaim attempt (node is still unbalanced). In that
4362 * case return the zone index of the previous kswapd reclaim cycle.
4363 */
4364static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
4365                                           enum zone_type prev_highest_zoneidx)
4366{
4367        enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
4368
4369        return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
4370}
4371
4372static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
4373                                unsigned int highest_zoneidx)
4374{
4375        long remaining = 0;
4376        DEFINE_WAIT(wait);
4377
4378        if (freezing(current) || kthread_should_stop())
4379                return;
4380
4381        prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
4382
4383        /*
4384         * Try to sleep for a short interval. Note that kcompactd will only be
4385         * woken if it is possible to sleep for a short interval. This is
4386         * deliberate on the assumption that if reclaim cannot keep an
4387         * eligible zone balanced that it's also unlikely that compaction will
4388         * succeed.
4389         */
4390        if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
4391                /*
4392                 * Compaction records what page blocks it recently failed to
4393                 * isolate pages from and skips them in the future scanning.
4394                 * When kswapd is going to sleep, it is reasonable to assume
4395                 * that pages and compaction may succeed so reset the cache.
4396                 */
4397                reset_isolation_suitable(pgdat);
4398
4399                /*
4400                 * We have freed the memory, now we should compact it to make
4401                 * allocation of the requested order possible.
4402                 */
4403                wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx);
4404
4405                remaining = schedule_timeout(HZ/10);
4406
4407                /*
4408                 * If woken prematurely then reset kswapd_highest_zoneidx and
4409                 * order. The values will either be from a wakeup request or
4410                 * the previous request that slept prematurely.
4411                 */
4412                if (remaining) {
4413                        WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
4414                                        kswapd_highest_zoneidx(pgdat,
4415                                                        highest_zoneidx));
4416
4417                        if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
4418                                WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
4419                }
4420
4421                finish_wait(&pgdat->kswapd_wait, &wait);
4422                prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
4423        }
4424
4425        /*
4426         * After a short sleep, check if it was a premature sleep. If not, then
4427         * go fully to sleep until explicitly woken up.
4428         */
4429        if (!remaining &&
4430            prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
4431                trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
4432
4433                /*
4434                 * vmstat counters are not perfectly accurate and the estimated
4435                 * value for counters such as NR_FREE_PAGES can deviate from the
4436                 * true value by nr_online_cpus * threshold. To avoid the zone
4437                 * watermarks being breached while under pressure, we reduce the
4438                 * per-cpu vmstat threshold while kswapd is awake and restore
4439                 * them before going back to sleep.
4440                 */
4441                set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
4442
4443                if (!kthread_should_stop())
4444                        schedule();
4445
4446                set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
4447        } else {
4448                if (remaining)
4449                        count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
4450                else
4451                        count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
4452        }
4453        finish_wait(&pgdat->kswapd_wait, &wait);
4454}
4455
4456/*
4457 * The background pageout daemon, started as a kernel thread
4458 * from the init process.
4459 *
4460 * This basically trickles out pages so that we have _some_
4461 * free memory available even if there is no other activity
4462 * that frees anything up. This is needed for things like routing
4463 * etc, where we otherwise might have all activity going on in
4464 * asynchronous contexts that cannot page things out.
4465 *
4466 * If there are applications that are active memory-allocators
4467 * (most normal use), this basically shouldn't matter.
4468 */
4469static int kswapd(void *p)
4470{
4471        unsigned int alloc_order, reclaim_order;
4472        unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
4473        pg_data_t *pgdat = (pg_data_t *)p;
4474        struct task_struct *tsk = current;
4475        const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
4476
4477        if (!cpumask_empty(cpumask))
4478                set_cpus_allowed_ptr(tsk, cpumask);
4479
4480        /*
4481         * Tell the memory management that we're a "memory allocator",
4482         * and that if we need more memory we should get access to it
4483         * regardless (see "__alloc_pages()"). "kswapd" should
4484         * never get caught in the normal page freeing logic.
4485         *
4486         * (Kswapd normally doesn't need memory anyway, but sometimes
4487         * you need a small amount of memory in order to be able to
4488         * page out something else, and this flag essentially protects
4489         * us from recursively trying to free more memory as we're
4490         * trying to free the first piece of memory in the first place).
4491         */
4492        tsk->flags |= PF_MEMALLOC | PF_KSWAPD;
4493        set_freezable();
4494
4495        WRITE_ONCE(pgdat->kswapd_order, 0);
4496        WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
4497        atomic_set(&pgdat->nr_writeback_throttled, 0);
4498        for ( ; ; ) {
4499                bool ret;
4500
4501                alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
4502                highest_zoneidx = kswapd_highest_zoneidx(pgdat,
4503                                                        highest_zoneidx);
4504
4505kswapd_try_sleep:
4506                kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
4507                                        highest_zoneidx);
4508
4509                /* Read the new order and highest_zoneidx */
4510                alloc_order = READ_ONCE(pgdat->kswapd_order);
4511                highest_zoneidx = kswapd_highest_zoneidx(pgdat,
4512                                                        highest_zoneidx);
4513                WRITE_ONCE(pgdat->kswapd_order, 0);
4514                WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
4515
4516                ret = try_to_freeze();
4517                if (kthread_should_stop())
4518                        break;
4519
4520                /*
4521                 * We can speed up thawing tasks if we don't call balance_pgdat
4522                 * after returning from the refrigerator
4523                 */
4524                if (ret)
4525                        continue;
4526
4527                /*
4528                 * Reclaim begins at the requested order but if a high-order
4529                 * reclaim fails then kswapd falls back to reclaiming for
4530                 * order-0. If that happens, kswapd will consider sleeping
4531                 * for the order it finished reclaiming at (reclaim_order)
4532                 * but kcompactd is woken to compact for the original
4533                 * request (alloc_order).
4534                 */
4535                trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx,
4536                                                alloc_order);
4537                reclaim_order = balance_pgdat(pgdat, alloc_order,
4538                                                highest_zoneidx);
4539                if (reclaim_order < alloc_order)
4540                        goto kswapd_try_sleep;
4541        }
4542
4543        tsk->flags &= ~(PF_MEMALLOC | PF_KSWAPD);
4544
4545        return 0;
4546}
4547
4548/*
4549 * A zone is low on free memory or too fragmented for high-order memory.  If
4550 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
4551 * pgdat.  It will wake up kcompactd after reclaiming memory.  If kswapd reclaim
4552 * has failed or is not needed, still wake up kcompactd if only compaction is
4553 * needed.
4554 */
4555void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
4556                   enum zone_type highest_zoneidx)
4557{
4558        pg_data_t *pgdat;
4559        enum zone_type curr_idx;
4560
4561        if (!managed_zone(zone))
4562                return;
4563
4564        if (!cpuset_zone_allowed(zone, gfp_flags))
4565                return;
4566
4567        pgdat = zone->zone_pgdat;
4568        curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
4569
4570        if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
4571                WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
4572
4573        if (READ_ONCE(pgdat->kswapd_order) < order)
4574                WRITE_ONCE(pgdat->kswapd_order, order);
4575
4576        if (!waitqueue_active(&pgdat->kswapd_wait))
4577                return;
4578
4579        /* Hopeless node, leave it to direct reclaim if possible */
4580        if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
4581            (pgdat_balanced(pgdat, order, highest_zoneidx) &&
4582             !pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
4583                /*
4584                 * There may be plenty of free memory available, but it's too
4585                 * fragmented for high-order allocations.  Wake up kcompactd
4586                 * and rely on compaction_suitable() to determine if it's
4587                 * needed.  If it fails, it will defer subsequent attempts to
4588                 * ratelimit its work.
4589                 */
4590                if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
4591                        wakeup_kcompactd(pgdat, order, highest_zoneidx);
4592                return;
4593        }
4594
4595        trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order,
4596                                      gfp_flags);
4597        wake_up_interruptible(&pgdat->kswapd_wait);
4598}
4599
4600#ifdef CONFIG_HIBERNATION
4601/*
4602 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
4603 * freed pages.
4604 *
4605 * Rather than trying to age LRUs the aim is to preserve the overall
4606 * LRU order by reclaiming preferentially
4607 * inactive > active > active referenced > active mapped
4608 */
4609unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
4610{
4611        struct scan_control sc = {
4612                .nr_to_reclaim = nr_to_reclaim,
4613                .gfp_mask = GFP_HIGHUSER_MOVABLE,
4614                .reclaim_idx = MAX_NR_ZONES - 1,
4615                .priority = DEF_PRIORITY,
4616                .may_writepage = 1,
4617                .may_unmap = 1,
4618                .may_swap = 1,
4619                .hibernation_mode = 1,
4620        };
4621        struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
4622        unsigned long nr_reclaimed;
4623        unsigned int noreclaim_flag;
4624
4625        fs_reclaim_acquire(sc.gfp_mask);
4626        noreclaim_flag = memalloc_noreclaim_save();
4627        set_task_reclaim_state(current, &sc.reclaim_state);
4628
4629        nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
4630
4631        set_task_reclaim_state(current, NULL);
4632        memalloc_noreclaim_restore(noreclaim_flag);
4633        fs_reclaim_release(sc.gfp_mask);
4634
4635        return nr_reclaimed;
4636}
4637#endif /* CONFIG_HIBERNATION */
4638
4639/*
4640 * This kswapd start function will be called by init and node-hot-add.
4641 */
4642void kswapd_run(int nid)
4643{
4644        pg_data_t *pgdat = NODE_DATA(nid);
4645
4646        if (pgdat->kswapd)
4647                return;
4648
4649        pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4650        if (IS_ERR(pgdat->kswapd)) {
4651                /* failure at boot is fatal */
4652                BUG_ON(system_state < SYSTEM_RUNNING);
4653                pr_err("Failed to start kswapd on node %d\n", nid);
4654                pgdat->kswapd = NULL;
4655        }
4656}
4657
4658/*
4659 * Called by memory hotplug when all memory in a node is offlined.  Caller must
4660 * be holding mem_hotplug_begin/done().
4661 */
4662void kswapd_stop(int nid)
4663{
4664        struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4665
4666        if (kswapd) {
4667                kthread_stop(kswapd);
4668                NODE_DATA(nid)->kswapd = NULL;
4669        }
4670}
4671
4672static int __init kswapd_init(void)
4673{
4674        int nid;
4675
4676        swap_setup();
4677        for_each_node_state(nid, N_MEMORY)
4678                kswapd_run(nid);
4679        return 0;
4680}
4681
4682module_init(kswapd_init)
4683
4684#ifdef CONFIG_NUMA
4685/*
4686 * Node reclaim mode
4687 *
4688 * If non-zero call node_reclaim when the number of free pages falls below
4689 * the watermarks.
4690 */
4691int node_reclaim_mode __read_mostly;
4692
4693/*
4694 * Priority for NODE_RECLAIM. This determines the fraction of pages
4695 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4696 * a zone.
4697 */
4698#define NODE_RECLAIM_PRIORITY 4
4699
4700/*
4701 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4702 * occur.
4703 */
4704int sysctl_min_unmapped_ratio = 1;
4705
4706/*
4707 * If the number of slab pages in a zone grows beyond this percentage then
4708 * slab reclaim needs to occur.
4709 */
4710int sysctl_min_slab_ratio = 5;
4711
4712static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4713{
4714        unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4715        unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4716                node_page_state(pgdat, NR_ACTIVE_FILE);
4717
4718        /*
4719         * It's possible for there to be more file mapped pages than
4720         * accounted for by the pages on the file LRU lists because
4721         * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4722         */
4723        return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4724}
4725
4726/* Work out how many page cache pages we can reclaim in this reclaim_mode */
4727static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4728{
4729        unsigned long nr_pagecache_reclaimable;
4730        unsigned long delta = 0;
4731
4732        /*
4733         * If RECLAIM_UNMAP is set, then all file pages are considered
4734         * potentially reclaimable. Otherwise, we have to worry about
4735         * pages like swapcache and node_unmapped_file_pages() provides
4736         * a better estimate
4737         */
4738        if (node_reclaim_mode & RECLAIM_UNMAP)
4739                nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4740        else
4741                nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4742
4743        /* If we can't clean pages, remove dirty pages from consideration */
4744        if (!(node_reclaim_mode & RECLAIM_WRITE))
4745                delta += node_page_state(pgdat, NR_FILE_DIRTY);
4746
4747        /* Watch for any possible underflows due to delta */
4748        if (unlikely(delta > nr_pagecache_reclaimable))
4749                delta = nr_pagecache_reclaimable;
4750
4751        return nr_pagecache_reclaimable - delta;
4752}
4753
4754/*
4755 * Try to free up some pages from this node through reclaim.
4756 */
4757static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4758{
4759        /* Minimum pages needed in order to stay on node */
4760        const unsigned long nr_pages = 1 << order;
4761        struct task_struct *p = current;
4762        unsigned int noreclaim_flag;
4763        struct scan_control sc = {
4764                .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4765                .gfp_mask = current_gfp_context(gfp_mask),
4766                .order = order,
4767                .priority = NODE_RECLAIM_PRIORITY,
4768                .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4769                .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4770                .may_swap = 1,
4771                .reclaim_idx = gfp_zone(gfp_mask),
4772        };
4773        unsigned long pflags;
4774
4775        trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4776                                           sc.gfp_mask);
4777
4778        cond_resched();
4779        psi_memstall_enter(&pflags);
4780        fs_reclaim_acquire(sc.gfp_mask);
4781        /*
4782         * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4783         */
4784        noreclaim_flag = memalloc_noreclaim_save();
4785        set_task_reclaim_state(p, &sc.reclaim_state);
4786
4787        if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages ||
4788            node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) > pgdat->min_slab_pages) {
4789                /*
4790                 * Free memory by calling shrink node with increasing
4791                 * priorities until we have enough memory freed.
4792                 */
4793                do {
4794                        shrink_node(pgdat, &sc);
4795                } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4796        }
4797
4798        set_task_reclaim_state(p, NULL);
4799        memalloc_noreclaim_restore(noreclaim_flag);
4800        fs_reclaim_release(sc.gfp_mask);
4801        psi_memstall_leave(&pflags);
4802
4803        trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4804
4805        return sc.nr_reclaimed >= nr_pages;
4806}
4807
4808int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4809{
4810        int ret;
4811
4812        /*
4813         * Node reclaim reclaims unmapped file backed pages and
4814         * slab pages if we are over the defined limits.
4815         *
4816         * A small portion of unmapped file backed pages is needed for
4817         * file I/O otherwise pages read by file I/O will be immediately
4818         * thrown out if the node is overallocated. So we do not reclaim
4819         * if less than a specified percentage of the node is used by
4820         * unmapped file backed pages.
4821         */
4822        if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4823            node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <=
4824            pgdat->min_slab_pages)
4825                return NODE_RECLAIM_FULL;
4826
4827        /*
4828         * Do not scan if the allocation should not be delayed.
4829         */
4830        if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4831                return NODE_RECLAIM_NOSCAN;
4832
4833        /*
4834         * Only run node reclaim on the local node or on nodes that do not
4835         * have associated processors. This will favor the local processor
4836         * over remote processors and spread off node memory allocations
4837         * as wide as possible.
4838         */
4839        if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4840                return NODE_RECLAIM_NOSCAN;
4841
4842        if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4843                return NODE_RECLAIM_NOSCAN;
4844
4845        ret = __node_reclaim(pgdat, gfp_mask, order);
4846        clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4847
4848        if (!ret)
4849                count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4850
4851        return ret;
4852}
4853#endif
4854
4855void check_move_unevictable_pages(struct pagevec *pvec)
4856{
4857        struct folio_batch fbatch;
4858        unsigned i;
4859
4860        folio_batch_init(&fbatch);
4861        for (i = 0; i < pvec->nr; i++) {
4862                struct page *page = pvec->pages[i];
4863
4864                if (PageTransTail(page))
4865                        continue;
4866                folio_batch_add(&fbatch, page_folio(page));
4867        }
4868        check_move_unevictable_folios(&fbatch);
4869}
4870EXPORT_SYMBOL_GPL(check_move_unevictable_pages);
4871
4872/**
4873 * check_move_unevictable_folios - Move evictable folios to appropriate zone
4874 * lru list
4875 * @fbatch: Batch of lru folios to check.
4876 *
4877 * Checks folios for evictability, if an evictable folio is in the unevictable
4878 * lru list, moves it to the appropriate evictable lru list. This function
4879 * should be only used for lru folios.
4880 */
4881void check_move_unevictable_folios(struct folio_batch *fbatch)
4882{
4883        struct lruvec *lruvec = NULL;
4884        int pgscanned = 0;
4885        int pgrescued = 0;
4886        int i;
4887
4888        for (i = 0; i < fbatch->nr; i++) {
4889                struct folio *folio = fbatch->folios[i];
4890                int nr_pages = folio_nr_pages(folio);
4891
4892                pgscanned += nr_pages;
4893
4894                /* block memcg migration while the folio moves between lrus */
4895                if (!folio_test_clear_lru(folio))
4896                        continue;
4897
4898                lruvec = folio_lruvec_relock_irq(folio, lruvec);
4899                if (folio_evictable(folio) && folio_test_unevictable(folio)) {
4900                        lruvec_del_folio(lruvec, folio);
4901                        folio_clear_unevictable(folio);
4902                        lruvec_add_folio(lruvec, folio);
4903                        pgrescued += nr_pages;
4904                }
4905                folio_set_lru(folio);
4906        }
4907
4908        if (lruvec) {
4909                __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4910                __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4911                unlock_page_lruvec_irq(lruvec);
4912        } else if (pgscanned) {
4913                count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4914        }
4915}
4916EXPORT_SYMBOL_GPL(check_move_unevictable_folios);
4917